SUPDrv.c@ 40904

Last change on this file since 40904 was 40904, checked in by vboxsync, 13 years ago
supdrvIOCtl: Reduce the number of return probes.
Property svn:eol-style set to `native` Property svn:keywords set to `Author Date Id Revision`
File size: 215.8 KB

Line
1	/* $Id: SUPDrv.c 40904 2012-04-13 14:47:07Z vboxsync $ */
2	/** @file
3	* VBoxDrv - The VirtualBox Support Driver - Common code.
4	*/
5
6	/*
7	* Copyright (C) 2006-2012 Oracle Corporation
8	*
9	* This file is part of VirtualBox Open Source Edition (OSE), as
10	* available from http://www.virtualbox.org. This file is free software;
11	* you can redistribute it and/or modify it under the terms of the GNU
12	* General Public License (GPL) as published by the Free Software
13	* Foundation, in version 2 as it comes in the "COPYING" file of the
14	* VirtualBox OSE distribution. VirtualBox OSE is distributed in the
15	* hope that it will be useful, but WITHOUT ANY WARRANTY of any kind.
16	*
17	* The contents of this file may alternatively be used under the terms
18	* of the Common Development and Distribution License Version 1.0
19	* (CDDL) only, as it comes in the "COPYING.CDDL" file of the
20	* VirtualBox OSE distribution, in which case the provisions of the
21	* CDDL are applicable instead of those of the GPL.
22	*
23	* You may elect to license modified versions of this file under the
24	* terms and conditions of either the GPL or the CDDL or both.
25	*/
26
27	/*******************************************************************************
28	* Header Files *
29	*******************************************************************************/
30	#define LOG_GROUP LOG_GROUP_SUP_DRV
31	#define SUPDRV_AGNOSTIC
32	#include "SUPDrvInternal.h"
33	#ifndef PAGE_SHIFT
34	# include <iprt/param.h>
35	#endif
36	#include <iprt/asm.h>
37	#include <iprt/asm-amd64-x86.h>
38	#include <iprt/asm-math.h>
39	#include <iprt/cpuset.h>
40	#include <iprt/handletable.h>
41	#include <iprt/mem.h>
42	#include <iprt/mp.h>
43	#include <iprt/power.h>
44	#include <iprt/process.h>
45	#include <iprt/semaphore.h>
46	#include <iprt/spinlock.h>
47	#include <iprt/thread.h>
48	#include <iprt/uuid.h>
49	#include <iprt/net.h>
50	#include <iprt/crc.h>
51	#include <iprt/string.h>
52	#include <iprt/timer.h>
53	#if defined(RT_OS_DARWIN) \|\| defined(RT_OS_SOLARIS) \|\| defined(RT_OS_FREEBSD)
54	# include <iprt/rand.h>
55	# include <iprt/path.h>
56	#endif
57	#include <iprt/x86.h>
58
59	#include <VBox/param.h>
60	#include <VBox/log.h>
61	#include <VBox/err.h>
62	#include <VBox/vmm/hwacc_svm.h>
63	#include <VBox/vmm/hwacc_vmx.h>
64
65	#if defined(RT_OS_SOLARIS) \|\| defined(RT_OS_DARWIN)
66	# include "dtrace/SUPDrv.h"
67	#else
68	# define VBOXDRV_SESSION_CREATE(pvSession, fUser) do { } while (0)
69	# define VBOXDRV_SESSION_CLOSE(pvSession) do { } while (0)
70	# define VBOXDRV_IOCTL_ENTRY(pvSession, uIOCtl, pvReqHdr) do { } while (0)
71	# define VBOXDRV_IOCTL_RETURN(pvSession, uIOCtl, pvReqHdr, rcRet, rcReq) do { } while (0)
72	#endif
73
74	/*
75	* Logging assignments:
76	* Log - useful stuff, like failures.
77	* LogFlow - program flow, except the really noisy bits.
78	* Log2 - Cleanup.
79	* Log3 - Loader flow noise.
80	* Log4 - Call VMMR0 flow noise.
81	* Log5 - Native yet-to-be-defined noise.
82	* Log6 - Native ioctl flow noise.
83	*
84	* Logging requires BUILD_TYPE=debug and possibly changes to the logger
85	* instantiation in log-vbox.c(pp).
86	*/
87
88
89	/*******************************************************************************
90	* Defined Constants And Macros *
91	*******************************************************************************/
92	/** The frequency by which we recalculate the u32UpdateHz and
93	* u32UpdateIntervalNS GIP members. The value must be a power of 2. */
94	#define GIP_UPDATEHZ_RECALC_FREQ 0x800
95
96	/** @def VBOX_SVN_REV
97	* The makefile should define this if it can. */
98	#ifndef VBOX_SVN_REV
99	# define VBOX_SVN_REV 0
100	#endif
101
102	#if 0 /* Don't start the GIP timers. Useful when debugging the IPRT timer code. */
103	# define DO_NOT_START_GIP
104	#endif
105
106
107	/*******************************************************************************
108	* Internal Functions *
109	*******************************************************************************/
110	static DECLCALLBACK(int) supdrvSessionObjHandleRetain(RTHANDLETABLE hHandleTable, void pvObj, void pvCtx, void *pvUser);
111	static DECLCALLBACK(void) supdrvSessionObjHandleDelete(RTHANDLETABLE hHandleTable, uint32_t h, void pvObj, void pvCtx, void *pvUser);
112	static int supdrvMemAdd(PSUPDRVMEMREF pMem, PSUPDRVSESSION pSession);
113	static int supdrvMemRelease(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, SUPDRVMEMREFTYPE eType);
114	static int supdrvIOCtl_LdrOpen(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDROPEN pReq);
115	static int supdrvIOCtl_LdrLoad(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRLOAD pReq);
116	static int supdrvIOCtl_LdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRFREE pReq);
117	static int supdrvIOCtl_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRGETSYMBOL pReq);
118	static int supdrvIDC_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVIDCREQGETSYM pReq);
119	static int supdrvLdrSetVMMR0EPs(PSUPDRVDEVEXT pDevExt, void pvVMMR0, void pvVMMR0EntryInt, void pvVMMR0EntryFast, void pvVMMR0EntryEx);
120	static void supdrvLdrUnsetVMMR0EPs(PSUPDRVDEVEXT pDevExt);
121	static int supdrvLdrAddUsage(PSUPDRVSESSION pSession, PSUPDRVLDRIMAGE pImage);
122	static void supdrvLdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVLDRIMAGE pImage);
123	DECLINLINE(int) supdrvLdrLock(PSUPDRVDEVEXT pDevExt);
124	DECLINLINE(int) supdrvLdrUnlock(PSUPDRVDEVEXT pDevExt);
125	static int supdrvIOCtl_CallServiceModule(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPCALLSERVICE pReq);
126	static int supdrvIOCtl_LoggerSettings(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLOGGERSETTINGS pReq);
127	static int supdrvGipCreate(PSUPDRVDEVEXT pDevExt);
128	static void supdrvGipDestroy(PSUPDRVDEVEXT pDevExt);
129	static DECLCALLBACK(void) supdrvGipSyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick);
130	static DECLCALLBACK(void) supdrvGipAsyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick);
131	static DECLCALLBACK(void) supdrvGipMpEvent(RTMPEVENT enmEvent, RTCPUID idCpu, void *pvUser);
132	static void supdrvGipInit(PSUPDRVDEVEXT pDevExt, PSUPGLOBALINFOPAGE pGip, RTHCPHYS HCPhys,
133	uint64_t u64NanoTS, unsigned uUpdateHz, unsigned cCpus);
134	static DECLCALLBACK(void) supdrvGipInitOnCpu(RTCPUID idCpu, void pvUser1, void pvUser2);
135	static void supdrvGipTerm(PSUPGLOBALINFOPAGE pGip);
136	static void supdrvGipUpdate(PSUPDRVDEVEXT pDevExt, uint64_t u64NanoTS, uint64_t u64TSC, RTCPUID idCpu, uint64_t iTick);
137	static void supdrvGipUpdatePerCpu(PSUPDRVDEVEXT pDevExt, uint64_t u64NanoTS, uint64_t u64TSC,
138	RTCPUID idCpu, uint8_t idApic, uint64_t iTick);
139	static void supdrvGipInitCpu(PSUPGLOBALINFOPAGE pGip, PSUPGIPCPU pCpu, uint64_t u64NanoTS);
140
141
142	/*******************************************************************************
143	* Global Variables *
144	*******************************************************************************/
145	DECLEXPORT(PSUPGLOBALINFOPAGE) g_pSUPGlobalInfoPage = NULL;
146
147	/**
148	* Array of the R0 SUP API.
149	*/
150	static SUPFUNC g_aFunctions[] =
151	{
152	/* SED: START */
153	/* name function */
154	/* Entries with absolute addresses determined at runtime, fixup
155	code makes ugly ASSUMPTIONS about the order here: */
156	{ "SUPR0AbsIs64bit", (void *)0 },
157	{ "SUPR0Abs64bitKernelCS", (void *)0 },
158	{ "SUPR0Abs64bitKernelSS", (void *)0 },
159	{ "SUPR0Abs64bitKernelDS", (void *)0 },
160	{ "SUPR0AbsKernelCS", (void *)0 },
161	{ "SUPR0AbsKernelSS", (void *)0 },
162	{ "SUPR0AbsKernelDS", (void *)0 },
163	{ "SUPR0AbsKernelES", (void *)0 },
164	{ "SUPR0AbsKernelFS", (void *)0 },
165	{ "SUPR0AbsKernelGS", (void *)0 },
166	/* Normal function pointers: */
167	{ "g_pSUPGlobalInfoPage", (void )&g_pSUPGlobalInfoPage }, / SED: DATA */
168	{ "SUPGetGIP", (void *)SUPGetGIP },
169	{ "SUPR0ComponentDeregisterFactory", (void *)SUPR0ComponentDeregisterFactory },
170	{ "SUPR0ComponentQueryFactory", (void *)SUPR0ComponentQueryFactory },
171	{ "SUPR0ComponentRegisterFactory", (void *)SUPR0ComponentRegisterFactory },
172	{ "SUPR0ContAlloc", (void *)SUPR0ContAlloc },
173	{ "SUPR0ContFree", (void *)SUPR0ContFree },
174	{ "SUPR0EnableVTx", (void *)SUPR0EnableVTx },
175	{ "SUPR0GetPagingMode", (void *)SUPR0GetPagingMode },
176	{ "SUPR0LockMem", (void *)SUPR0LockMem },
177	{ "SUPR0LowAlloc", (void *)SUPR0LowAlloc },
178	{ "SUPR0LowFree", (void *)SUPR0LowFree },
179	{ "SUPR0MemAlloc", (void *)SUPR0MemAlloc },
180	{ "SUPR0MemFree", (void *)SUPR0MemFree },
181	{ "SUPR0MemGetPhys", (void *)SUPR0MemGetPhys },
182	{ "SUPR0ObjAddRef", (void *)SUPR0ObjAddRef },
183	{ "SUPR0ObjAddRefEx", (void *)SUPR0ObjAddRefEx },
184	{ "SUPR0ObjRegister", (void *)SUPR0ObjRegister },
185	{ "SUPR0ObjRelease", (void *)SUPR0ObjRelease },
186	{ "SUPR0ObjVerifyAccess", (void *)SUPR0ObjVerifyAccess },
187	{ "SUPR0PageAllocEx", (void *)SUPR0PageAllocEx },
188	{ "SUPR0PageFree", (void *)SUPR0PageFree },
189	{ "SUPR0Printf", (void *)SUPR0Printf },
190	{ "SUPR0TracerDeregisterDrv", (void *)SUPR0TracerDeregisterDrv },
191	{ "SUPR0TracerDeregisterImpl", (void *)SUPR0TracerDeregisterImpl },
192	{ "SUPR0TracerFireProbe", (void *)SUPR0TracerFireProbe },
193	{ "SUPR0TracerRegisterDrv", (void *)SUPR0TracerRegisterDrv },
194	{ "SUPR0TracerRegisterImpl", (void *)SUPR0TracerRegisterImpl },
195	{ "SUPR0TracerRegisterModule", (void *)SUPR0TracerRegisterModule },
196	{ "SUPR0UnlockMem", (void *)SUPR0UnlockMem },
197	{ "SUPSemEventClose", (void *)SUPSemEventClose },
198	{ "SUPSemEventCreate", (void *)SUPSemEventCreate },
199	{ "SUPSemEventGetResolution", (void *)SUPSemEventGetResolution },
200	{ "SUPSemEventMultiClose", (void *)SUPSemEventMultiClose },
201	{ "SUPSemEventMultiCreate", (void *)SUPSemEventMultiCreate },
202	{ "SUPSemEventMultiGetResolution", (void *)SUPSemEventMultiGetResolution },
203	{ "SUPSemEventMultiReset", (void *)SUPSemEventMultiReset },
204	{ "SUPSemEventMultiSignal", (void *)SUPSemEventMultiSignal },
205	{ "SUPSemEventMultiWait", (void *)SUPSemEventMultiWait },
206	{ "SUPSemEventMultiWaitNoResume", (void *)SUPSemEventMultiWaitNoResume },
207	{ "SUPSemEventMultiWaitNsAbsIntr", (void *)SUPSemEventMultiWaitNsAbsIntr },
208	{ "SUPSemEventMultiWaitNsRelIntr", (void *)SUPSemEventMultiWaitNsRelIntr },
209	{ "SUPSemEventSignal", (void *)SUPSemEventSignal },
210	{ "SUPSemEventWait", (void *)SUPSemEventWait },
211	{ "SUPSemEventWaitNoResume", (void *)SUPSemEventWaitNoResume },
212	{ "SUPSemEventWaitNsAbsIntr", (void *)SUPSemEventWaitNsAbsIntr },
213	{ "SUPSemEventWaitNsRelIntr", (void *)SUPSemEventWaitNsRelIntr },
214
215	{ "RTAssertAreQuiet", (void *)RTAssertAreQuiet },
216	{ "RTAssertMayPanic", (void *)RTAssertMayPanic },
217	{ "RTAssertMsg1", (void *)RTAssertMsg1 },
218	{ "RTAssertMsg2AddV", (void *)RTAssertMsg2AddV },
219	{ "RTAssertMsg2V", (void *)RTAssertMsg2V },
220	{ "RTAssertSetMayPanic", (void *)RTAssertSetMayPanic },
221	{ "RTAssertSetQuiet", (void *)RTAssertSetQuiet },
222	{ "RTCrc32", (void *)RTCrc32 },
223	{ "RTCrc32Finish", (void *)RTCrc32Finish },
224	{ "RTCrc32Process", (void *)RTCrc32Process },
225	{ "RTCrc32Start", (void *)RTCrc32Start },
226	{ "RTErrConvertFromErrno", (void *)RTErrConvertFromErrno },
227	{ "RTErrConvertToErrno", (void *)RTErrConvertToErrno },
228	{ "RTHandleTableAllocWithCtx", (void *)RTHandleTableAllocWithCtx },
229	{ "RTHandleTableCreate", (void *)RTHandleTableCreate },
230	{ "RTHandleTableCreateEx", (void *)RTHandleTableCreateEx },
231	{ "RTHandleTableDestroy", (void *)RTHandleTableDestroy },
232	{ "RTHandleTableFreeWithCtx", (void *)RTHandleTableFreeWithCtx },
233	{ "RTHandleTableLookupWithCtx", (void *)RTHandleTableLookupWithCtx },
234	{ "RTLogDefaultInstance", (void *)RTLogDefaultInstance },
235	{ "RTLogLoggerExV", (void *)RTLogLoggerExV },
236	{ "RTLogPrintfV", (void *)RTLogPrintfV },
237	{ "RTLogRelDefaultInstance", (void *)RTLogRelDefaultInstance },
238	{ "RTLogSetDefaultInstanceThread", (void *)RTLogSetDefaultInstanceThread },
239	{ "RTMemAllocExTag", (void *)RTMemAllocExTag },
240	{ "RTMemAllocTag", (void *)RTMemAllocTag },
241	{ "RTMemAllocVarTag", (void *)RTMemAllocVarTag },
242	{ "RTMemAllocZTag", (void *)RTMemAllocZTag },
243	{ "RTMemAllocZVarTag", (void *)RTMemAllocZVarTag },
244	{ "RTMemDupExTag", (void *)RTMemDupExTag },
245	{ "RTMemDupTag", (void *)RTMemDupTag },
246	{ "RTMemFree", (void *)RTMemFree },
247	{ "RTMemFreeEx", (void *)RTMemFreeEx },
248	{ "RTMemReallocTag", (void *)RTMemReallocTag },
249	{ "RTMpCpuId", (void *)RTMpCpuId },
250	{ "RTMpCpuIdFromSetIndex", (void *)RTMpCpuIdFromSetIndex },
251	{ "RTMpCpuIdToSetIndex", (void *)RTMpCpuIdToSetIndex },
252	{ "RTMpGetArraySize", (void *)RTMpGetArraySize },
253	{ "RTMpGetCount", (void *)RTMpGetCount },
254	{ "RTMpGetMaxCpuId", (void *)RTMpGetMaxCpuId },
255	{ "RTMpGetOnlineCount", (void *)RTMpGetOnlineCount },
256	{ "RTMpGetOnlineSet", (void *)RTMpGetOnlineSet },
257	{ "RTMpGetSet", (void *)RTMpGetSet },
258	{ "RTMpIsCpuOnline", (void *)RTMpIsCpuOnline },
259	{ "RTMpIsCpuPossible", (void *)RTMpIsCpuPossible },
260	{ "RTMpIsCpuWorkPending", (void *)RTMpIsCpuWorkPending },
261	{ "RTMpNotificationDeregister", (void *)RTMpNotificationDeregister },
262	{ "RTMpNotificationRegister", (void *)RTMpNotificationRegister },
263	{ "RTMpOnAll", (void *)RTMpOnAll },
264	{ "RTMpOnOthers", (void *)RTMpOnOthers },
265	{ "RTMpOnSpecific", (void *)RTMpOnSpecific },
266	{ "RTMpPokeCpu", (void *)RTMpPokeCpu },
267	{ "RTNetIPv4AddDataChecksum", (void *)RTNetIPv4AddDataChecksum },
268	{ "RTNetIPv4AddTCPChecksum", (void *)RTNetIPv4AddTCPChecksum },
269	{ "RTNetIPv4AddUDPChecksum", (void *)RTNetIPv4AddUDPChecksum },
270	{ "RTNetIPv4FinalizeChecksum", (void *)RTNetIPv4FinalizeChecksum },
271	{ "RTNetIPv4HdrChecksum", (void *)RTNetIPv4HdrChecksum },
272	{ "RTNetIPv4IsDHCPValid", (void *)RTNetIPv4IsDHCPValid },
273	{ "RTNetIPv4IsHdrValid", (void *)RTNetIPv4IsHdrValid },
274	{ "RTNetIPv4IsTCPSizeValid", (void *)RTNetIPv4IsTCPSizeValid },
275	{ "RTNetIPv4IsTCPValid", (void *)RTNetIPv4IsTCPValid },
276	{ "RTNetIPv4IsUDPSizeValid", (void *)RTNetIPv4IsUDPSizeValid },
277	{ "RTNetIPv4IsUDPValid", (void *)RTNetIPv4IsUDPValid },
278	{ "RTNetIPv4PseudoChecksum", (void *)RTNetIPv4PseudoChecksum },
279	{ "RTNetIPv4PseudoChecksumBits", (void *)RTNetIPv4PseudoChecksumBits },
280	{ "RTNetIPv4TCPChecksum", (void *)RTNetIPv4TCPChecksum },
281	{ "RTNetIPv4UDPChecksum", (void *)RTNetIPv4UDPChecksum },
282	{ "RTNetIPv6PseudoChecksum", (void *)RTNetIPv6PseudoChecksum },
283	{ "RTNetIPv6PseudoChecksumBits", (void *)RTNetIPv6PseudoChecksumBits },
284	{ "RTNetIPv6PseudoChecksumEx", (void *)RTNetIPv6PseudoChecksumEx },
285	{ "RTNetTCPChecksum", (void *)RTNetTCPChecksum },
286	{ "RTNetUDPChecksum", (void *)RTNetUDPChecksum },
287	{ "RTPowerNotificationDeregister", (void *)RTPowerNotificationDeregister },
288	{ "RTPowerNotificationRegister", (void *)RTPowerNotificationRegister },
289	{ "RTProcSelf", (void *)RTProcSelf },
290	{ "RTR0AssertPanicSystem", (void *)RTR0AssertPanicSystem },
291	{ "RTR0MemAreKrnlAndUsrDifferent", (void *)RTR0MemAreKrnlAndUsrDifferent },
292	{ "RTR0MemKernelIsValidAddr", (void *)RTR0MemKernelIsValidAddr },
293	{ "RTR0MemObjAddress", (void *)RTR0MemObjAddress },
294	{ "RTR0MemObjAddressR3", (void *)RTR0MemObjAddressR3 },
295	{ "RTR0MemObjAllocContTag", (void *)RTR0MemObjAllocContTag },
296	{ "RTR0MemObjAllocLowTag", (void *)RTR0MemObjAllocLowTag },
297	{ "RTR0MemObjAllocPageTag", (void *)RTR0MemObjAllocPageTag },
298	{ "RTR0MemObjAllocPhysExTag", (void *)RTR0MemObjAllocPhysExTag },
299	{ "RTR0MemObjAllocPhysNCTag", (void *)RTR0MemObjAllocPhysNCTag },
300	{ "RTR0MemObjAllocPhysTag", (void *)RTR0MemObjAllocPhysTag },
301	{ "RTR0MemObjEnterPhysTag", (void *)RTR0MemObjEnterPhysTag },
302	{ "RTR0MemObjFree", (void *)RTR0MemObjFree },
303	{ "RTR0MemObjGetPagePhysAddr", (void *)RTR0MemObjGetPagePhysAddr },
304	{ "RTR0MemObjIsMapping", (void *)RTR0MemObjIsMapping },
305	{ "RTR0MemObjLockUserTag", (void *)RTR0MemObjLockUserTag },
306	{ "RTR0MemObjMapKernelExTag", (void *)RTR0MemObjMapKernelExTag },
307	{ "RTR0MemObjMapKernelTag", (void *)RTR0MemObjMapKernelTag },
308	{ "RTR0MemObjMapUserTag", (void *)RTR0MemObjMapUserTag },
309	{ "RTR0MemObjProtect", (void *)RTR0MemObjProtect },
310	{ "RTR0MemObjSize", (void *)RTR0MemObjSize },
311	{ "RTR0MemUserCopyFrom", (void *)RTR0MemUserCopyFrom },
312	{ "RTR0MemUserCopyTo", (void *)RTR0MemUserCopyTo },
313	{ "RTR0MemUserIsValidAddr", (void *)RTR0MemUserIsValidAddr },
314	{ "RTR0ProcHandleSelf", (void *)RTR0ProcHandleSelf },
315	{ "RTSemEventCreate", (void *)RTSemEventCreate },
316	{ "RTSemEventDestroy", (void *)RTSemEventDestroy },
317	{ "RTSemEventGetResolution", (void *)RTSemEventGetResolution },
318	{ "RTSemEventMultiCreate", (void *)RTSemEventMultiCreate },
319	{ "RTSemEventMultiDestroy", (void *)RTSemEventMultiDestroy },
320	{ "RTSemEventMultiGetResolution", (void *)RTSemEventMultiGetResolution },
321	{ "RTSemEventMultiReset", (void *)RTSemEventMultiReset },
322	{ "RTSemEventMultiSignal", (void *)RTSemEventMultiSignal },
323	{ "RTSemEventMultiWait", (void *)RTSemEventMultiWait },
324	{ "RTSemEventMultiWaitEx", (void *)RTSemEventMultiWaitEx },
325	{ "RTSemEventMultiWaitExDebug", (void *)RTSemEventMultiWaitExDebug },
326	{ "RTSemEventMultiWaitNoResume", (void *)RTSemEventMultiWaitNoResume },
327	{ "RTSemEventSignal", (void *)RTSemEventSignal },
328	{ "RTSemEventWait", (void *)RTSemEventWait },
329	{ "RTSemEventWaitEx", (void *)RTSemEventWaitEx },
330	{ "RTSemEventWaitExDebug", (void *)RTSemEventWaitExDebug },
331	{ "RTSemEventWaitNoResume", (void *)RTSemEventWaitNoResume },
332	{ "RTSemFastMutexCreate", (void *)RTSemFastMutexCreate },
333	{ "RTSemFastMutexDestroy", (void *)RTSemFastMutexDestroy },
334	{ "RTSemFastMutexRelease", (void *)RTSemFastMutexRelease },
335	{ "RTSemFastMutexRequest", (void *)RTSemFastMutexRequest },
336	{ "RTSemMutexCreate", (void *)RTSemMutexCreate },
337	{ "RTSemMutexDestroy", (void *)RTSemMutexDestroy },
338	{ "RTSemMutexRelease", (void *)RTSemMutexRelease },
339	{ "RTSemMutexRequest", (void *)RTSemMutexRequest },
340	{ "RTSemMutexRequestDebug", (void *)RTSemMutexRequestDebug },
341	{ "RTSemMutexRequestNoResume", (void *)RTSemMutexRequestNoResume },
342	{ "RTSemMutexRequestNoResumeDebug", (void *)RTSemMutexRequestNoResumeDebug },
343	{ "RTSpinlockAcquire", (void *)RTSpinlockAcquire },
344	{ "RTSpinlockCreate", (void *)RTSpinlockCreate },
345	{ "RTSpinlockDestroy", (void *)RTSpinlockDestroy },
346	{ "RTSpinlockRelease", (void *)RTSpinlockRelease },
347	{ "RTSpinlockReleaseNoInts", (void *)RTSpinlockReleaseNoInts },
348	{ "RTStrCopy", (void *)RTStrCopy },
349	{ "RTStrDupTag", (void *)RTStrDupTag },
350	{ "RTStrFormat", (void *)RTStrFormat },
351	{ "RTStrFormatNumber", (void *)RTStrFormatNumber },
352	{ "RTStrFormatTypeDeregister", (void *)RTStrFormatTypeDeregister },
353	{ "RTStrFormatTypeRegister", (void *)RTStrFormatTypeRegister },
354	{ "RTStrFormatTypeSetUser", (void *)RTStrFormatTypeSetUser },
355	{ "RTStrFormatV", (void *)RTStrFormatV },
356	{ "RTStrFree", (void *)RTStrFree },
357	{ "RTStrNCmp", (void *)RTStrNCmp },
358	{ "RTStrPrintf", (void *)RTStrPrintf },
359	{ "RTStrPrintfEx", (void *)RTStrPrintfEx },
360	{ "RTStrPrintfExV", (void *)RTStrPrintfExV },
361	{ "RTStrPrintfV", (void *)RTStrPrintfV },
362	{ "RTThreadCreate", (void *)RTThreadCreate },
363	{ "RTThreadGetName", (void *)RTThreadGetName },
364	{ "RTThreadGetNative", (void *)RTThreadGetNative },
365	{ "RTThreadGetType", (void *)RTThreadGetType },
366	{ "RTThreadIsInInterrupt", (void *)RTThreadIsInInterrupt },
367	{ "RTThreadNativeSelf", (void *)RTThreadNativeSelf },
368	{ "RTThreadPreemptDisable", (void *)RTThreadPreemptDisable },
369	{ "RTThreadPreemptIsEnabled", (void *)RTThreadPreemptIsEnabled },
370	{ "RTThreadPreemptIsPending", (void *)RTThreadPreemptIsPending },
371	{ "RTThreadPreemptIsPendingTrusty", (void *)RTThreadPreemptIsPendingTrusty },
372	{ "RTThreadPreemptIsPossible", (void *)RTThreadPreemptIsPossible },
373	{ "RTThreadPreemptRestore", (void *)RTThreadPreemptRestore },
374	{ "RTThreadSelf", (void *)RTThreadSelf },
375	{ "RTThreadSelfName", (void *)RTThreadSelfName },
376	{ "RTThreadSleep", (void *)RTThreadSleep },
377	{ "RTThreadUserReset", (void *)RTThreadUserReset },
378	{ "RTThreadUserSignal", (void *)RTThreadUserSignal },
379	{ "RTThreadUserWait", (void *)RTThreadUserWait },
380	{ "RTThreadUserWaitNoResume", (void *)RTThreadUserWaitNoResume },
381	{ "RTThreadWait", (void *)RTThreadWait },
382	{ "RTThreadWaitNoResume", (void *)RTThreadWaitNoResume },
383	{ "RTThreadYield", (void *)RTThreadYield },
384	{ "RTTimeMilliTS", (void *)RTTimeMilliTS },
385	{ "RTTimeNanoTS", (void *)RTTimeNanoTS },
386	{ "RTTimeNow", (void *)RTTimeNow },
387	{ "RTTimerCanDoHighResolution", (void *)RTTimerCanDoHighResolution },
388	{ "RTTimerChangeInterval", (void *)RTTimerChangeInterval },
389	{ "RTTimerCreate", (void *)RTTimerCreate },
390	{ "RTTimerCreateEx", (void *)RTTimerCreateEx },
391	{ "RTTimerDestroy", (void *)RTTimerDestroy },
392	{ "RTTimerGetSystemGranularity", (void *)RTTimerGetSystemGranularity },
393	{ "RTTimerReleaseSystemGranularity", (void *)RTTimerReleaseSystemGranularity },
394	{ "RTTimerRequestSystemGranularity", (void *)RTTimerRequestSystemGranularity },
395	{ "RTTimerStart", (void *)RTTimerStart },
396	{ "RTTimerStop", (void *)RTTimerStop },
397	{ "RTTimeSystemMilliTS", (void *)RTTimeSystemMilliTS },
398	{ "RTTimeSystemNanoTS", (void *)RTTimeSystemNanoTS },
399	{ "RTUuidCompare", (void *)RTUuidCompare },
400	{ "RTUuidCompareStr", (void *)RTUuidCompareStr },
401	{ "RTUuidFromStr", (void *)RTUuidFromStr },
402	/* SED: END */
403	};
404
405	#if defined(RT_OS_DARWIN) \|\| defined(RT_OS_SOLARIS) \|\| defined(RT_OS_FREEBSD)
406	/**
407	* Drag in the rest of IRPT since we share it with the
408	* rest of the kernel modules on darwin.
409	*/
410	PFNRT g_apfnVBoxDrvIPRTDeps[] =
411	{
412	/* VBoxNetAdp */
413	(PFNRT)RTRandBytes,
414	/* VBoxUSB */
415	(PFNRT)RTPathStripFilename,
416	NULL
417	};
418	#endif /* RT_OS_DARWIN \|\| RT_OS_SOLARIS \|\| RT_OS_SOLARIS */
419
420
421	/**
422	* Initializes the device extentsion structure.
423	*
424	* @returns IPRT status code.
425	* @param pDevExt The device extension to initialize.
426	* @param cbSession The size of the session structure. The size of
427	* SUPDRVSESSION may be smaller when SUPDRV_AGNOSTIC is
428	* defined because we're skipping the OS specific members
429	* then.
430	*/
431	int VBOXCALL supdrvInitDevExt(PSUPDRVDEVEXT pDevExt, size_t cbSession)
432	{
433	int rc;
434
435	#ifdef SUPDRV_WITH_RELEASE_LOGGER
436	/*
437	* Create the release log.
438	*/
439	static const char * const s_apszGroups[] = VBOX_LOGGROUP_NAMES;
440	PRTLOGGER pRelLogger;
441	rc = RTLogCreate(&pRelLogger, 0 /* fFlags */, "all",
442	"VBOX_RELEASE_LOG", RT_ELEMENTS(s_apszGroups), s_apszGroups, RTLOGDEST_STDOUT \| RTLOGDEST_DEBUGGER, NULL);
443	if (RT_SUCCESS(rc))
444	RTLogRelSetDefaultInstance(pRelLogger);
445	/** @todo Add native hook for getting logger config parameters and setting
446	* them. On linux we should use the module parameter stuff... */
447	#endif
448
449	/*
450	* Initialize it.
451	*/
452	memset(pDevExt, 0, sizeof(*pDevExt));
453	rc = RTSpinlockCreate(&pDevExt->Spinlock, RTSPINLOCK_FLAGS_INTERRUPT_UNSAFE, "SUPDrvDevExt");
454	if (RT_SUCCESS(rc))
455	{
456	rc = RTSpinlockCreate(&pDevExt->hGipSpinlock, RTSPINLOCK_FLAGS_INTERRUPT_SAFE, "SUPDrvGip");
457	if (RT_SUCCESS(rc))
458	{
459	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
460	rc = RTSemMutexCreate(&pDevExt->mtxLdr);
461	#else
462	rc = RTSemFastMutexCreate(&pDevExt->mtxLdr);
463	#endif
464	if (RT_SUCCESS(rc))
465	{
466	rc = RTSemFastMutexCreate(&pDevExt->mtxComponentFactory);
467	if (RT_SUCCESS(rc))
468	{
469	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
470	rc = RTSemMutexCreate(&pDevExt->mtxGip);
471	#else
472	rc = RTSemFastMutexCreate(&pDevExt->mtxGip);
473	#endif
474	if (RT_SUCCESS(rc))
475	{
476	rc = supdrvGipCreate(pDevExt);
477	if (RT_SUCCESS(rc))
478	{
479	rc = supdrvTracerInit(pDevExt);
480	if (RT_SUCCESS(rc))
481	{
482	pDevExt->pLdrInitImage = NULL;
483	pDevExt->hLdrInitThread = NIL_RTNATIVETHREAD;
484	pDevExt->u32Cookie = BIRD; /** @todo make this random? */
485	pDevExt->cbSession = (uint32_t)cbSession;
486
487	/*
488	* Fixup the absolute symbols.
489	*
490	* Because of the table indexing assumptions we'll have a little #ifdef orgy
491	* here rather than distributing this to OS specific files. At least for now.
492	*/
493	#ifdef RT_OS_DARWIN
494	# if ARCH_BITS == 32
495	if (SUPR0GetPagingMode() >= SUPPAGINGMODE_AMD64)
496	{
497	g_aFunctions[0].pfn = (void )1; / SUPR0AbsIs64bit */
498	g_aFunctions[1].pfn = (void )0x80; / SUPR0Abs64bitKernelCS - KERNEL64_CS, seg.h */
499	g_aFunctions[2].pfn = (void )0x88; / SUPR0Abs64bitKernelSS - KERNEL64_SS, seg.h */
500	g_aFunctions[3].pfn = (void )0x88; / SUPR0Abs64bitKernelDS - KERNEL64_SS, seg.h */
501	}
502	else
503	g_aFunctions[0].pfn = g_aFunctions[1].pfn = g_aFunctions[2].pfn = g_aFunctions[4].pfn = (void *)0;
504	g_aFunctions[4].pfn = (void )0x08; / SUPR0AbsKernelCS - KERNEL_CS, seg.h */
505	g_aFunctions[5].pfn = (void )0x10; / SUPR0AbsKernelSS - KERNEL_DS, seg.h */
506	g_aFunctions[6].pfn = (void )0x10; / SUPR0AbsKernelDS - KERNEL_DS, seg.h */
507	g_aFunctions[7].pfn = (void )0x10; / SUPR0AbsKernelES - KERNEL_DS, seg.h */
508	g_aFunctions[8].pfn = (void )0x10; / SUPR0AbsKernelFS - KERNEL_DS, seg.h */
509	g_aFunctions[9].pfn = (void )0x48; / SUPR0AbsKernelGS - CPU_DATA_GS, seg.h */
510	# else /* 64-bit darwin: */
511	g_aFunctions[0].pfn = (void )1; / SUPR0AbsIs64bit */
512	g_aFunctions[1].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0Abs64bitKernelCS */
513	g_aFunctions[2].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0Abs64bitKernelSS */
514	g_aFunctions[3].pfn = (void )0; / SUPR0Abs64bitKernelDS */
515	g_aFunctions[4].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0AbsKernelCS */
516	g_aFunctions[5].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0AbsKernelSS */
517	g_aFunctions[6].pfn = (void )0; / SUPR0AbsKernelDS */
518	g_aFunctions[7].pfn = (void )0; / SUPR0AbsKernelES */
519	g_aFunctions[8].pfn = (void )0; / SUPR0AbsKernelFS */
520	g_aFunctions[9].pfn = (void )0; / SUPR0AbsKernelGS */
521
522	# endif
523	#else /* !RT_OS_DARWIN */
524	# if ARCH_BITS == 64
525	g_aFunctions[0].pfn = (void )1; / SUPR0AbsIs64bit */
526	g_aFunctions[1].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0Abs64bitKernelCS */
527	g_aFunctions[2].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0Abs64bitKernelSS */
528	g_aFunctions[3].pfn = (void )(uintptr_t)ASMGetDS(); / SUPR0Abs64bitKernelDS */
529	# else
530	g_aFunctions[0].pfn = g_aFunctions[1].pfn = g_aFunctions[2].pfn = g_aFunctions[4].pfn = (void *)0;
531	# endif
532	g_aFunctions[4].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0AbsKernelCS */
533	g_aFunctions[5].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0AbsKernelSS */
534	g_aFunctions[6].pfn = (void )(uintptr_t)ASMGetDS(); / SUPR0AbsKernelDS */
535	g_aFunctions[7].pfn = (void )(uintptr_t)ASMGetES(); / SUPR0AbsKernelES */
536	g_aFunctions[8].pfn = (void )(uintptr_t)ASMGetFS(); / SUPR0AbsKernelFS */
537	g_aFunctions[9].pfn = (void )(uintptr_t)ASMGetGS(); / SUPR0AbsKernelGS */
538	#endif /* !RT_OS_DARWIN */
539	return VINF_SUCCESS;
540	}
541
542	supdrvGipDestroy(pDevExt);
543	}
544
545	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
546	RTSemMutexDestroy(pDevExt->mtxGip);
547	pDevExt->mtxGip = NIL_RTSEMMUTEX;
548	#else
549	RTSemFastMutexDestroy(pDevExt->mtxGip);
550	pDevExt->mtxGip = NIL_RTSEMFASTMUTEX;
551	#endif
552	}
553	RTSemFastMutexDestroy(pDevExt->mtxComponentFactory);
554	pDevExt->mtxComponentFactory = NIL_RTSEMFASTMUTEX;
555	}
556	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
557	RTSemMutexDestroy(pDevExt->mtxLdr);
558	pDevExt->mtxLdr = NIL_RTSEMMUTEX;
559	#else
560	RTSemFastMutexDestroy(pDevExt->mtxLdr);
561	pDevExt->mtxLdr = NIL_RTSEMFASTMUTEX;
562	#endif
563	}
564	RTSpinlockDestroy(pDevExt->hGipSpinlock);
565	pDevExt->hGipSpinlock = NIL_RTSPINLOCK;
566	}
567	RTSpinlockDestroy(pDevExt->Spinlock);
568	pDevExt->Spinlock = NIL_RTSPINLOCK;
569	}
570	#ifdef SUPDRV_WITH_RELEASE_LOGGER
571	RTLogDestroy(RTLogRelSetDefaultInstance(NULL));
572	RTLogDestroy(RTLogSetDefaultInstance(NULL));
573	#endif
574
575	return rc;
576	}
577
578
579	/**
580	* Delete the device extension (e.g. cleanup members).
581	*
582	* @param pDevExt The device extension to delete.
583	*/
584	void VBOXCALL supdrvDeleteDevExt(PSUPDRVDEVEXT pDevExt)
585	{
586	PSUPDRVOBJ pObj;
587	PSUPDRVUSAGE pUsage;
588
589	/*
590	* Kill mutexes and spinlocks.
591	*/
592	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
593	RTSemMutexDestroy(pDevExt->mtxGip);
594	pDevExt->mtxGip = NIL_RTSEMMUTEX;
595	#else
596	RTSemFastMutexDestroy(pDevExt->mtxGip);
597	pDevExt->mtxGip = NIL_RTSEMFASTMUTEX;
598	#endif
599	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
600	RTSemMutexDestroy(pDevExt->mtxLdr);
601	pDevExt->mtxLdr = NIL_RTSEMMUTEX;
602	#else
603	RTSemFastMutexDestroy(pDevExt->mtxLdr);
604	pDevExt->mtxLdr = NIL_RTSEMFASTMUTEX;
605	#endif
606	RTSpinlockDestroy(pDevExt->Spinlock);
607	pDevExt->Spinlock = NIL_RTSPINLOCK;
608	RTSemFastMutexDestroy(pDevExt->mtxComponentFactory);
609	pDevExt->mtxComponentFactory = NIL_RTSEMFASTMUTEX;
610
611	/*
612	* Free lists.
613	*/
614	/* objects. */
615	pObj = pDevExt->pObjs;
616	Assert(!pObj); /* (can trigger on forced unloads) */
617	pDevExt->pObjs = NULL;
618	while (pObj)
619	{
620	void *pvFree = pObj;
621	pObj = pObj->pNext;
622	RTMemFree(pvFree);
623	}
624
625	/* usage records. */
626	pUsage = pDevExt->pUsageFree;
627	pDevExt->pUsageFree = NULL;
628	while (pUsage)
629	{
630	void *pvFree = pUsage;
631	pUsage = pUsage->pNext;
632	RTMemFree(pvFree);
633	}
634
635	/* kill the GIP. */
636	supdrvGipDestroy(pDevExt);
637	RTSpinlockDestroy(pDevExt->hGipSpinlock);
638	pDevExt->hGipSpinlock = NIL_RTSPINLOCK;
639
640	supdrvTracerTerm(pDevExt);
641
642	#ifdef SUPDRV_WITH_RELEASE_LOGGER
643	/* destroy the loggers. */
644	RTLogDestroy(RTLogRelSetDefaultInstance(NULL));
645	RTLogDestroy(RTLogSetDefaultInstance(NULL));
646	#endif
647	}
648
649
650	/**
651	* Create session.
652	*
653	* @returns IPRT status code.
654	* @param pDevExt Device extension.
655	* @param fUser Flag indicating whether this is a user or kernel session.
656	* @param ppSession Where to store the pointer to the session data.
657	*/
658	int VBOXCALL supdrvCreateSession(PSUPDRVDEVEXT pDevExt, bool fUser, PSUPDRVSESSION *ppSession)
659	{
660	/*
661	* Allocate memory for the session data.
662	*/
663	int rc;
664	PSUPDRVSESSION pSession = *ppSession = (PSUPDRVSESSION)RTMemAllocZ(pDevExt->cbSession);
665	if (pSession)
666	{
667	/* Initialize session data. */
668	rc = RTSpinlockCreate(&pSession->Spinlock, RTSPINLOCK_FLAGS_INTERRUPT_UNSAFE, "SUPDrvSession");
669	if (!rc)
670	{
671	rc = RTHandleTableCreateEx(&pSession->hHandleTable,
672	RTHANDLETABLE_FLAGS_LOCKED \| RTHANDLETABLE_FLAGS_CONTEXT,
673	1 /uBase/, 32768 /cMax/, supdrvSessionObjHandleRetain, pSession);
674	if (RT_SUCCESS(rc))
675	{
676	Assert(pSession->Spinlock != NIL_RTSPINLOCK);
677	pSession->pDevExt = pDevExt;
678	pSession->u32Cookie = BIRD_INV;
679	/*pSession->pLdrUsage = NULL;
680	pSession->pVM = NULL;
681	pSession->pUsage = NULL;
682	pSession->pGip = NULL;
683	pSession->fGipReferenced = false;
684	pSession->Bundle.cUsed = 0; */
685	pSession->Uid = NIL_RTUID;
686	pSession->Gid = NIL_RTGID;
687	if (fUser)
688	{
689	pSession->Process = RTProcSelf();
690	pSession->R0Process = RTR0ProcHandleSelf();
691	}
692	else
693	{
694	pSession->Process = NIL_RTPROCESS;
695	pSession->R0Process = NIL_RTR0PROCESS;
696	}
697	/pSession->uTracerData = 0;/
698	pSession->hTracerCaller = NIL_RTNATIVETHREAD;
699
700	VBOXDRV_SESSION_CREATE(pSession, fUser);
701	LogFlow(("Created session %p initial cookie=%#x\n", pSession, pSession->u32Cookie));
702	return VINF_SUCCESS;
703	}
704
705	RTSpinlockDestroy(pSession->Spinlock);
706	}
707	RTMemFree(pSession);
708	*ppSession = NULL;
709	Log(("Failed to create spinlock, rc=%d!\n", rc));
710	}
711	else
712	rc = VERR_NO_MEMORY;
713
714	return rc;
715	}
716
717
718	/**
719	* Shared code for cleaning up a session.
720	*
721	* @param pDevExt Device extension.
722	* @param pSession Session data.
723	* This data will be freed by this routine.
724	*/
725	void VBOXCALL supdrvCloseSession(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
726	{
727	VBOXDRV_SESSION_CLOSE(pSession);
728
729	/*
730	* Cleanup the session first.
731	*/
732	supdrvCleanupSession(pDevExt, pSession);
733
734	/*
735	* Free the rest of the session stuff.
736	*/
737	RTSpinlockDestroy(pSession->Spinlock);
738	pSession->Spinlock = NIL_RTSPINLOCK;
739	pSession->pDevExt = NULL;
740	RTMemFree(pSession);
741	LogFlow(("supdrvCloseSession: returns\n"));
742	}
743
744
745	/**
746	* Shared code for cleaning up a session (but not quite freeing it).
747	*
748	* This is primarily intended for MAC OS X where we have to clean up the memory
749	* stuff before the file handle is closed.
750	*
751	* @param pDevExt Device extension.
752	* @param pSession Session data.
753	* This data will be freed by this routine.
754	*/
755	void VBOXCALL supdrvCleanupSession(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
756	{
757	int rc;
758	PSUPDRVBUNDLE pBundle;
759	LogFlow(("supdrvCleanupSession: pSession=%p\n", pSession));
760
761	/*
762	* Remove logger instances related to this session.
763	*/
764	RTLogSetDefaultInstanceThread(NULL, (uintptr_t)pSession);
765
766	/*
767	* Destroy the handle table.
768	*/
769	rc = RTHandleTableDestroy(pSession->hHandleTable, supdrvSessionObjHandleDelete, pSession);
770	AssertRC(rc);
771	pSession->hHandleTable = NIL_RTHANDLETABLE;
772
773	/*
774	* Release object references made in this session.
775	* In theory there should be noone racing us in this session.
776	*/
777	Log2(("release objects - start\n"));
778	if (pSession->pUsage)
779	{
780	PSUPDRVUSAGE pUsage;
781	RTSpinlockAcquire(pDevExt->Spinlock);
782
783	while ((pUsage = pSession->pUsage) != NULL)
784	{
785	PSUPDRVOBJ pObj = pUsage->pObj;
786	pSession->pUsage = pUsage->pNext;
787
788	AssertMsg(pUsage->cUsage >= 1 && pObj->cUsage >= pUsage->cUsage, ("glob %d; sess %d\n", pObj->cUsage, pUsage->cUsage));
789	if (pUsage->cUsage < pObj->cUsage)
790	{
791	pObj->cUsage -= pUsage->cUsage;
792	RTSpinlockRelease(pDevExt->Spinlock);
793	}
794	else
795	{
796	/* Destroy the object and free the record. */
797	if (pDevExt->pObjs == pObj)
798	pDevExt->pObjs = pObj->pNext;
799	else
800	{
801	PSUPDRVOBJ pObjPrev;
802	for (pObjPrev = pDevExt->pObjs; pObjPrev; pObjPrev = pObjPrev->pNext)
803	if (pObjPrev->pNext == pObj)
804	{
805	pObjPrev->pNext = pObj->pNext;
806	break;
807	}
808	Assert(pObjPrev);
809	}
810	RTSpinlockRelease(pDevExt->Spinlock);
811
812	Log(("supdrvCleanupSession: destroying %p/%d (%p/%p) cpid=%RTproc pid=%RTproc dtor=%p\n",
813	pObj, pObj->enmType, pObj->pvUser1, pObj->pvUser2, pObj->CreatorProcess, RTProcSelf(), pObj->pfnDestructor));
814	if (pObj->pfnDestructor)
815	pObj->pfnDestructor(pObj, pObj->pvUser1, pObj->pvUser2);
816	RTMemFree(pObj);
817	}
818
819	/* free it and continue. */
820	RTMemFree(pUsage);
821
822	RTSpinlockAcquire(pDevExt->Spinlock);
823	}
824
825	RTSpinlockRelease(pDevExt->Spinlock);
826	AssertMsg(!pSession->pUsage, ("Some buster reregistered an object during desturction!\n"));
827	}
828	Log2(("release objects - done\n"));
829
830	/*
831	* Do tracer cleanups related to this session.
832	*/
833	Log2(("release tracer stuff - start\n"));
834	supdrvTracerCleanupSession(pDevExt, pSession);
835	Log2(("release tracer stuff - end\n"));
836
837	/*
838	* Release memory allocated in the session.
839	*
840	* We do not serialize this as we assume that the application will
841	* not allocated memory while closing the file handle object.
842	*/
843	Log2(("freeing memory:\n"));
844	pBundle = &pSession->Bundle;
845	while (pBundle)
846	{
847	PSUPDRVBUNDLE pToFree;
848	unsigned i;
849
850	/*
851	* Check and unlock all entries in the bundle.
852	*/
853	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
854	{
855	if (pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ)
856	{
857	Log2(("eType=%d pvR0=%p pvR3=%p cb=%ld\n", pBundle->aMem[i].eType, RTR0MemObjAddress(pBundle->aMem[i].MemObj),
858	(void *)RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3), (long)RTR0MemObjSize(pBundle->aMem[i].MemObj)));
859	if (pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ)
860	{
861	rc = RTR0MemObjFree(pBundle->aMem[i].MapObjR3, false);
862	AssertRC(rc); /** @todo figure out how to handle this. */
863	pBundle->aMem[i].MapObjR3 = NIL_RTR0MEMOBJ;
864	}
865	rc = RTR0MemObjFree(pBundle->aMem[i].MemObj, true /* fFreeMappings */);
866	AssertRC(rc); /** @todo figure out how to handle this. */
867	pBundle->aMem[i].MemObj = NIL_RTR0MEMOBJ;
868	pBundle->aMem[i].eType = MEMREF_TYPE_UNUSED;
869	}
870	}
871
872	/*
873	* Advance and free previous bundle.
874	*/
875	pToFree = pBundle;
876	pBundle = pBundle->pNext;
877
878	pToFree->pNext = NULL;
879	pToFree->cUsed = 0;
880	if (pToFree != &pSession->Bundle)
881	RTMemFree(pToFree);
882	}
883	Log2(("freeing memory - done\n"));
884
885	/*
886	* Deregister component factories.
887	*/
888	RTSemFastMutexRequest(pDevExt->mtxComponentFactory);
889	Log2(("deregistering component factories:\n"));
890	if (pDevExt->pComponentFactoryHead)
891	{
892	PSUPDRVFACTORYREG pPrev = NULL;
893	PSUPDRVFACTORYREG pCur = pDevExt->pComponentFactoryHead;
894	while (pCur)
895	{
896	if (pCur->pSession == pSession)
897	{
898	/* unlink it */
899	PSUPDRVFACTORYREG pNext = pCur->pNext;
900	if (pPrev)
901	pPrev->pNext = pNext;
902	else
903	pDevExt->pComponentFactoryHead = pNext;
904
905	/* free it */
906	pCur->pNext = NULL;
907	pCur->pSession = NULL;
908	pCur->pFactory = NULL;
909	RTMemFree(pCur);
910
911	/* next */
912	pCur = pNext;
913	}
914	else
915	{
916	/* next */
917	pPrev = pCur;
918	pCur = pCur->pNext;
919	}
920	}
921	}
922	RTSemFastMutexRelease(pDevExt->mtxComponentFactory);
923	Log2(("deregistering component factories - done\n"));
924
925	/*
926	* Loaded images needs to be dereferenced and possibly freed up.
927	*/
928	supdrvLdrLock(pDevExt);
929	Log2(("freeing images:\n"));
930	if (pSession->pLdrUsage)
931	{
932	PSUPDRVLDRUSAGE pUsage = pSession->pLdrUsage;
933	pSession->pLdrUsage = NULL;
934	while (pUsage)
935	{
936	void *pvFree = pUsage;
937	PSUPDRVLDRIMAGE pImage = pUsage->pImage;
938	if (pImage->cUsage > pUsage->cUsage)
939	pImage->cUsage -= pUsage->cUsage;
940	else
941	supdrvLdrFree(pDevExt, pImage);
942	pUsage->pImage = NULL;
943	pUsage = pUsage->pNext;
944	RTMemFree(pvFree);
945	}
946	}
947	supdrvLdrUnlock(pDevExt);
948	Log2(("freeing images - done\n"));
949
950	/*
951	* Unmap the GIP.
952	*/
953	Log2(("umapping GIP:\n"));
954	if (pSession->GipMapObjR3 != NIL_RTR0MEMOBJ)
955	{
956	SUPR0GipUnmap(pSession);
957	pSession->fGipReferenced = 0;
958	}
959	Log2(("umapping GIP - done\n"));
960	}
961
962
963	/**
964	* RTHandleTableDestroy callback used by supdrvCleanupSession.
965	*
966	* @returns IPRT status code, see SUPR0ObjAddRef.
967	* @param hHandleTable The handle table handle. Ignored.
968	* @param pvObj The object pointer.
969	* @param pvCtx Context, the handle type. Ignored.
970	* @param pvUser Session pointer.
971	*/
972	static DECLCALLBACK(int) supdrvSessionObjHandleRetain(RTHANDLETABLE hHandleTable, void pvObj, void pvCtx, void *pvUser)
973	{
974	NOREF(pvCtx);
975	NOREF(hHandleTable);
976	return SUPR0ObjAddRefEx(pvObj, (PSUPDRVSESSION)pvUser, true /fNoBlocking/);
977	}
978
979
980	/**
981	* RTHandleTableDestroy callback used by supdrvCleanupSession.
982	*
983	* @param hHandleTable The handle table handle. Ignored.
984	* @param h The handle value. Ignored.
985	* @param pvObj The object pointer.
986	* @param pvCtx Context, the handle type. Ignored.
987	* @param pvUser Session pointer.
988	*/
989	static DECLCALLBACK(void) supdrvSessionObjHandleDelete(RTHANDLETABLE hHandleTable, uint32_t h, void pvObj, void pvCtx, void *pvUser)
990	{
991	NOREF(pvCtx);
992	NOREF(h);
993	NOREF(hHandleTable);
994	SUPR0ObjRelease(pvObj, (PSUPDRVSESSION)pvUser);
995	}
996
997
998	/**
999	* Fast path I/O Control worker.
1000	*
1001	* @returns VBox status code that should be passed down to ring-3 unchanged.
1002	* @param uIOCtl Function number.
1003	* @param idCpu VMCPU id.
1004	* @param pDevExt Device extention.
1005	* @param pSession Session data.
1006	*/
1007	int VBOXCALL supdrvIOCtlFast(uintptr_t uIOCtl, VMCPUID idCpu, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
1008	{
1009	/*
1010	* We check the two prereqs after doing this only to allow the compiler to optimize things better.
1011	*/
1012	if (RT_LIKELY( RT_VALID_PTR(pSession)
1013	&& pSession->pVM
1014	&& pDevExt->pfnVMMR0EntryFast))
1015	{
1016	switch (uIOCtl)
1017	{
1018	case SUP_IOCTL_FAST_DO_RAW_RUN:
1019	pDevExt->pfnVMMR0EntryFast(pSession->pVM, idCpu, SUP_VMMR0_DO_RAW_RUN);
1020	break;
1021	case SUP_IOCTL_FAST_DO_HWACC_RUN:
1022	pDevExt->pfnVMMR0EntryFast(pSession->pVM, idCpu, SUP_VMMR0_DO_HWACC_RUN);
1023	break;
1024	case SUP_IOCTL_FAST_DO_NOP:
1025	pDevExt->pfnVMMR0EntryFast(pSession->pVM, idCpu, SUP_VMMR0_DO_NOP);
1026	break;
1027	default:
1028	return VERR_INTERNAL_ERROR;
1029	}
1030	return VINF_SUCCESS;
1031	}
1032	return VERR_INTERNAL_ERROR;
1033	}
1034
1035
1036	/**
1037	* Helper for supdrvIOCtl. Check if pszStr contains any character of pszChars.
1038	* We would use strpbrk here if this function would be contained in the RedHat kABI white
1039	* list, see http://www.kerneldrivers.org/RHEL5.
1040	*
1041	* @returns 1 if pszStr does contain any character of pszChars, 0 otherwise.
1042	* @param pszStr String to check
1043	* @param pszChars Character set
1044	*/
1045	static int supdrvCheckInvalidChar(const char pszStr, const char pszChars)
1046	{
1047	int chCur;
1048	while ((chCur = *pszStr++) != '\0')
1049	{
1050	int ch;
1051	const char *psz = pszChars;
1052	while ((ch = *psz++) != '\0')
1053	if (ch == chCur)
1054	return 1;
1055
1056	}
1057	return 0;
1058	}
1059
1060
1061
1062	/**
1063	* I/O Control inner worker (tracing reasons).
1064	*
1065	* @returns IPRT status code.
1066	* @retval VERR_INVALID_PARAMETER if the request is invalid.
1067	*
1068	* @param uIOCtl Function number.
1069	* @param pDevExt Device extention.
1070	* @param pSession Session data.
1071	* @param pReqHdr The request header.
1072	*/
1073	static int supdrvIOCtlInner(uintptr_t uIOCtl, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPREQHDR pReqHdr)
1074	{
1075	/*
1076	* Validation macros
1077	*/
1078	#define REQ_CHECK_SIZES_EX(Name, cbInExpect, cbOutExpect) \
1079	do { \
1080	if (RT_UNLIKELY(pReqHdr->cbIn != (cbInExpect) \|\| pReqHdr->cbOut != (cbOutExpect))) \
1081	{ \
1082	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbIn=%ld expected %ld. cbOut=%ld expected %ld.\n", \
1083	(long)pReqHdr->cbIn, (long)(cbInExpect), (long)pReqHdr->cbOut, (long)(cbOutExpect))); \
1084	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
1085	} \
1086	} while (0)
1087
1088	#define REQ_CHECK_SIZES(Name) REQ_CHECK_SIZES_EX(Name, Name ## _SIZE_IN, Name ## _SIZE_OUT)
1089
1090	#define REQ_CHECK_SIZE_IN(Name, cbInExpect) \
1091	do { \
1092	if (RT_UNLIKELY(pReqHdr->cbIn != (cbInExpect))) \
1093	{ \
1094	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbIn=%ld expected %ld.\n", \
1095	(long)pReqHdr->cbIn, (long)(cbInExpect))); \
1096	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
1097	} \
1098	} while (0)
1099
1100	#define REQ_CHECK_SIZE_OUT(Name, cbOutExpect) \
1101	do { \
1102	if (RT_UNLIKELY(pReqHdr->cbOut != (cbOutExpect))) \
1103	{ \
1104	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbOut=%ld expected %ld.\n", \
1105	(long)pReqHdr->cbOut, (long)(cbOutExpect))); \
1106	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
1107	} \
1108	} while (0)
1109
1110	#define REQ_CHECK_EXPR(Name, expr) \
1111	do { \
1112	if (RT_UNLIKELY(!(expr))) \
1113	{ \
1114	OSDBGPRINT(( #Name ": %s\n", #expr)); \
1115	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
1116	} \
1117	} while (0)
1118
1119	#define REQ_CHECK_EXPR_FMT(expr, fmt) \
1120	do { \
1121	if (RT_UNLIKELY(!(expr))) \
1122	{ \
1123	OSDBGPRINT( fmt ); \
1124	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
1125	} \
1126	} while (0)
1127
1128	/*
1129	* The switch.
1130	*/
1131	switch (SUP_CTL_CODE_NO_SIZE(uIOCtl))
1132	{
1133	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_COOKIE):
1134	{
1135	PSUPCOOKIE pReq = (PSUPCOOKIE)pReqHdr;
1136	REQ_CHECK_SIZES(SUP_IOCTL_COOKIE);
1137	if (strncmp(pReq->u.In.szMagic, SUPCOOKIE_MAGIC, sizeof(pReq->u.In.szMagic)))
1138	{
1139	OSDBGPRINT(("SUP_IOCTL_COOKIE: invalid magic %.16s\n", pReq->u.In.szMagic));
1140	pReq->Hdr.rc = VERR_INVALID_MAGIC;
1141	return 0;
1142	}
1143
1144	#if 0
1145	/*
1146	* Call out to the OS specific code and let it do permission checks on the
1147	* client process.
1148	*/
1149	if (!supdrvOSValidateClientProcess(pDevExt, pSession))
1150	{
1151	pReq->u.Out.u32Cookie = 0xffffffff;
1152	pReq->u.Out.u32SessionCookie = 0xffffffff;
1153	pReq->u.Out.u32SessionVersion = 0xffffffff;
1154	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1155	pReq->u.Out.pSession = NULL;
1156	pReq->u.Out.cFunctions = 0;
1157	pReq->Hdr.rc = VERR_PERMISSION_DENIED;
1158	return 0;
1159	}
1160	#endif
1161
1162	/*
1163	* Match the version.
1164	* The current logic is very simple, match the major interface version.
1165	*/
1166	if ( pReq->u.In.u32MinVersion > SUPDRV_IOC_VERSION
1167	\|\| (pReq->u.In.u32MinVersion & 0xffff0000) != (SUPDRV_IOC_VERSION & 0xffff0000))
1168	{
1169	OSDBGPRINT(("SUP_IOCTL_COOKIE: Version mismatch. Requested: %#x Min: %#x Current: %#x\n",
1170	pReq->u.In.u32ReqVersion, pReq->u.In.u32MinVersion, SUPDRV_IOC_VERSION));
1171	pReq->u.Out.u32Cookie = 0xffffffff;
1172	pReq->u.Out.u32SessionCookie = 0xffffffff;
1173	pReq->u.Out.u32SessionVersion = 0xffffffff;
1174	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1175	pReq->u.Out.pSession = NULL;
1176	pReq->u.Out.cFunctions = 0;
1177	pReq->Hdr.rc = VERR_VERSION_MISMATCH;
1178	return 0;
1179	}
1180
1181	/*
1182	* Fill in return data and be gone.
1183	* N.B. The first one to change SUPDRV_IOC_VERSION shall makes sure that
1184	* u32SessionVersion <= u32ReqVersion!
1185	*/
1186	/** @todo Somehow validate the client and negotiate a secure cookie... */
1187	pReq->u.Out.u32Cookie = pDevExt->u32Cookie;
1188	pReq->u.Out.u32SessionCookie = pSession->u32Cookie;
1189	pReq->u.Out.u32SessionVersion = SUPDRV_IOC_VERSION;
1190	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1191	pReq->u.Out.pSession = pSession;
1192	pReq->u.Out.cFunctions = sizeof(g_aFunctions) / sizeof(g_aFunctions[0]);
1193	pReq->Hdr.rc = VINF_SUCCESS;
1194	return 0;
1195	}
1196
1197	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_QUERY_FUNCS(0)):
1198	{
1199	/* validate */
1200	PSUPQUERYFUNCS pReq = (PSUPQUERYFUNCS)pReqHdr;
1201	REQ_CHECK_SIZES_EX(SUP_IOCTL_QUERY_FUNCS, SUP_IOCTL_QUERY_FUNCS_SIZE_IN, SUP_IOCTL_QUERY_FUNCS_SIZE_OUT(RT_ELEMENTS(g_aFunctions)));
1202
1203	/* execute */
1204	pReq->u.Out.cFunctions = RT_ELEMENTS(g_aFunctions);
1205	memcpy(&pReq->u.Out.aFunctions[0], g_aFunctions, sizeof(g_aFunctions));
1206	pReq->Hdr.rc = VINF_SUCCESS;
1207	return 0;
1208	}
1209
1210	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_LOCK):
1211	{
1212	/* validate */
1213	PSUPPAGELOCK pReq = (PSUPPAGELOCK)pReqHdr;
1214	REQ_CHECK_SIZE_IN(SUP_IOCTL_PAGE_LOCK, SUP_IOCTL_PAGE_LOCK_SIZE_IN);
1215	REQ_CHECK_SIZE_OUT(SUP_IOCTL_PAGE_LOCK, SUP_IOCTL_PAGE_LOCK_SIZE_OUT(pReq->u.In.cPages));
1216	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_LOCK, pReq->u.In.cPages > 0);
1217	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_LOCK, pReq->u.In.pvR3 >= PAGE_SIZE);
1218
1219	/* execute */
1220	pReq->Hdr.rc = SUPR0LockMem(pSession, pReq->u.In.pvR3, pReq->u.In.cPages, &pReq->u.Out.aPages[0]);
1221	if (RT_FAILURE(pReq->Hdr.rc))
1222	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1223	return 0;
1224	}
1225
1226	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_UNLOCK):
1227	{
1228	/* validate */
1229	PSUPPAGEUNLOCK pReq = (PSUPPAGEUNLOCK)pReqHdr;
1230	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_UNLOCK);
1231
1232	/* execute */
1233	pReq->Hdr.rc = SUPR0UnlockMem(pSession, pReq->u.In.pvR3);
1234	return 0;
1235	}
1236
1237	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CONT_ALLOC):
1238	{
1239	/* validate */
1240	PSUPCONTALLOC pReq = (PSUPCONTALLOC)pReqHdr;
1241	REQ_CHECK_SIZES(SUP_IOCTL_CONT_ALLOC);
1242
1243	/* execute */
1244	pReq->Hdr.rc = SUPR0ContAlloc(pSession, pReq->u.In.cPages, &pReq->u.Out.pvR0, &pReq->u.Out.pvR3, &pReq->u.Out.HCPhys);
1245	if (RT_FAILURE(pReq->Hdr.rc))
1246	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1247	return 0;
1248	}
1249
1250	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CONT_FREE):
1251	{
1252	/* validate */
1253	PSUPCONTFREE pReq = (PSUPCONTFREE)pReqHdr;
1254	REQ_CHECK_SIZES(SUP_IOCTL_CONT_FREE);
1255
1256	/* execute */
1257	pReq->Hdr.rc = SUPR0ContFree(pSession, (RTHCUINTPTR)pReq->u.In.pvR3);
1258	return 0;
1259	}
1260
1261	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_OPEN):
1262	{
1263	/* validate */
1264	PSUPLDROPEN pReq = (PSUPLDROPEN)pReqHdr;
1265	REQ_CHECK_SIZES(SUP_IOCTL_LDR_OPEN);
1266	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageWithTabs > 0);
1267	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageWithTabs < 16*_1M);
1268	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageBits > 0);
1269	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageBits > 0);
1270	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageBits < pReq->u.In.cbImageWithTabs);
1271	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.szName[0]);
1272	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, RTStrEnd(pReq->u.In.szName, sizeof(pReq->u.In.szName)));
1273	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, !supdrvCheckInvalidChar(pReq->u.In.szName, ";:()[]{}/\\\|&*%#@!~`\"'"));
1274	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, RTStrEnd(pReq->u.In.szFilename, sizeof(pReq->u.In.szFilename)));
1275
1276	/* execute */
1277	pReq->Hdr.rc = supdrvIOCtl_LdrOpen(pDevExt, pSession, pReq);
1278	return 0;
1279	}
1280
1281	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_LOAD):
1282	{
1283	/* validate */
1284	PSUPLDRLOAD pReq = (PSUPLDRLOAD)pReqHdr;
1285	REQ_CHECK_EXPR(Name, pReq->Hdr.cbIn >= sizeof(*pReq));
1286	REQ_CHECK_SIZES_EX(SUP_IOCTL_LDR_LOAD, SUP_IOCTL_LDR_LOAD_SIZE_IN(pReq->u.In.cbImageWithTabs), SUP_IOCTL_LDR_LOAD_SIZE_OUT);
1287	REQ_CHECK_EXPR(SUP_IOCTL_LDR_LOAD, pReq->u.In.cSymbols <= 16384);
1288	REQ_CHECK_EXPR_FMT( !pReq->u.In.cSymbols
1289	\|\| ( pReq->u.In.offSymbols < pReq->u.In.cbImageWithTabs
1290	&& pReq->u.In.offSymbols + pReq->u.In.cSymbols * sizeof(SUPLDRSYM) <= pReq->u.In.cbImageWithTabs),
1291	("SUP_IOCTL_LDR_LOAD: offSymbols=%#lx cSymbols=%#lx cbImageWithTabs=%#lx\n", (long)pReq->u.In.offSymbols,
1292	(long)pReq->u.In.cSymbols, (long)pReq->u.In.cbImageWithTabs));
1293	REQ_CHECK_EXPR_FMT( !pReq->u.In.cbStrTab
1294	\|\| ( pReq->u.In.offStrTab < pReq->u.In.cbImageWithTabs
1295	&& pReq->u.In.offStrTab + pReq->u.In.cbStrTab <= pReq->u.In.cbImageWithTabs
1296	&& pReq->u.In.cbStrTab <= pReq->u.In.cbImageWithTabs),
1297	("SUP_IOCTL_LDR_LOAD: offStrTab=%#lx cbStrTab=%#lx cbImageWithTabs=%#lx\n", (long)pReq->u.In.offStrTab,
1298	(long)pReq->u.In.cbStrTab, (long)pReq->u.In.cbImageWithTabs));
1299
1300	if (pReq->u.In.cSymbols)
1301	{
1302	uint32_t i;
1303	PSUPLDRSYM paSyms = (PSUPLDRSYM)&pReq->u.In.abImage[pReq->u.In.offSymbols];
1304	for (i = 0; i < pReq->u.In.cSymbols; i++)
1305	{
1306	REQ_CHECK_EXPR_FMT(paSyms[i].offSymbol < pReq->u.In.cbImageWithTabs,
1307	("SUP_IOCTL_LDR_LOAD: sym #%ld: symb off %#lx (max=%#lx)\n", (long)i, (long)paSyms[i].offSymbol, (long)pReq->u.In.cbImageWithTabs));
1308	REQ_CHECK_EXPR_FMT(paSyms[i].offName < pReq->u.In.cbStrTab,
1309	("SUP_IOCTL_LDR_LOAD: sym #%ld: name off %#lx (max=%#lx)\n", (long)i, (long)paSyms[i].offName, (long)pReq->u.In.cbImageWithTabs));
1310	REQ_CHECK_EXPR_FMT(RTStrEnd((char const *)&pReq->u.In.abImage[pReq->u.In.offStrTab + paSyms[i].offName],
1311	pReq->u.In.cbStrTab - paSyms[i].offName),
1312	("SUP_IOCTL_LDR_LOAD: sym #%ld: unterminated name! (%#lx / %#lx)\n", (long)i, (long)paSyms[i].offName, (long)pReq->u.In.cbImageWithTabs));
1313	}
1314	}
1315
1316	/* execute */
1317	pReq->Hdr.rc = supdrvIOCtl_LdrLoad(pDevExt, pSession, pReq);
1318	return 0;
1319	}
1320
1321	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_FREE):
1322	{
1323	/* validate */
1324	PSUPLDRFREE pReq = (PSUPLDRFREE)pReqHdr;
1325	REQ_CHECK_SIZES(SUP_IOCTL_LDR_FREE);
1326
1327	/* execute */
1328	pReq->Hdr.rc = supdrvIOCtl_LdrFree(pDevExt, pSession, pReq);
1329	return 0;
1330	}
1331
1332	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_GET_SYMBOL):
1333	{
1334	/* validate */
1335	PSUPLDRGETSYMBOL pReq = (PSUPLDRGETSYMBOL)pReqHdr;
1336	REQ_CHECK_SIZES(SUP_IOCTL_LDR_GET_SYMBOL);
1337	REQ_CHECK_EXPR(SUP_IOCTL_LDR_GET_SYMBOL, RTStrEnd(pReq->u.In.szSymbol, sizeof(pReq->u.In.szSymbol)));
1338
1339	/* execute */
1340	pReq->Hdr.rc = supdrvIOCtl_LdrGetSymbol(pDevExt, pSession, pReq);
1341	return 0;
1342	}
1343
1344	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CALL_VMMR0(0)):
1345	{
1346	/* validate */
1347	PSUPCALLVMMR0 pReq = (PSUPCALLVMMR0)pReqHdr;
1348	Log4(("SUP_IOCTL_CALL_VMMR0: op=%u in=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1349	pReq->u.In.uOperation, pReq->Hdr.cbIn, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1350
1351	if (pReq->Hdr.cbIn == SUP_IOCTL_CALL_VMMR0_SIZE(0))
1352	{
1353	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_VMMR0, SUP_IOCTL_CALL_VMMR0_SIZE_IN(0), SUP_IOCTL_CALL_VMMR0_SIZE_OUT(0));
1354
1355	/* execute */
1356	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
1357	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pReq->u.In.pVMR0, pReq->u.In.idCpu, pReq->u.In.uOperation, NULL, pReq->u.In.u64Arg, pSession);
1358	else
1359	pReq->Hdr.rc = VERR_WRONG_ORDER;
1360	}
1361	else
1362	{
1363	PSUPVMMR0REQHDR pVMMReq = (PSUPVMMR0REQHDR)&pReq->abReqPkt[0];
1364	REQ_CHECK_EXPR_FMT(pReq->Hdr.cbIn >= SUP_IOCTL_CALL_VMMR0_SIZE(sizeof(SUPVMMR0REQHDR)),
1365	("SUP_IOCTL_CALL_VMMR0: cbIn=%#x < %#lx\n", pReq->Hdr.cbIn, SUP_IOCTL_CALL_VMMR0_SIZE(sizeof(SUPVMMR0REQHDR))));
1366	REQ_CHECK_EXPR(SUP_IOCTL_CALL_VMMR0, pVMMReq->u32Magic == SUPVMMR0REQHDR_MAGIC);
1367	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_VMMR0, SUP_IOCTL_CALL_VMMR0_SIZE_IN(pVMMReq->cbReq), SUP_IOCTL_CALL_VMMR0_SIZE_OUT(pVMMReq->cbReq));
1368
1369	/* execute */
1370	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
1371	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pReq->u.In.pVMR0, pReq->u.In.idCpu, pReq->u.In.uOperation, pVMMReq, pReq->u.In.u64Arg, pSession);
1372	else
1373	pReq->Hdr.rc = VERR_WRONG_ORDER;
1374	}
1375
1376	if ( RT_FAILURE(pReq->Hdr.rc)
1377	&& pReq->Hdr.rc != VERR_INTERRUPTED
1378	&& pReq->Hdr.rc != VERR_TIMEOUT)
1379	Log(("SUP_IOCTL_CALL_VMMR0: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1380	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1381	else
1382	Log4(("SUP_IOCTL_CALL_VMMR0: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1383	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1384	return 0;
1385	}
1386
1387	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CALL_VMMR0_BIG):
1388	{
1389	/* validate */
1390	PSUPCALLVMMR0 pReq = (PSUPCALLVMMR0)pReqHdr;
1391	PSUPVMMR0REQHDR pVMMReq;
1392	Log4(("SUP_IOCTL_CALL_VMMR0_BIG: op=%u in=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1393	pReq->u.In.uOperation, pReq->Hdr.cbIn, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1394
1395	pVMMReq = (PSUPVMMR0REQHDR)&pReq->abReqPkt[0];
1396	REQ_CHECK_EXPR_FMT(pReq->Hdr.cbIn >= SUP_IOCTL_CALL_VMMR0_BIG_SIZE(sizeof(SUPVMMR0REQHDR)),
1397	("SUP_IOCTL_CALL_VMMR0_BIG: cbIn=%#x < %#lx\n", pReq->Hdr.cbIn, SUP_IOCTL_CALL_VMMR0_BIG_SIZE(sizeof(SUPVMMR0REQHDR))));
1398	REQ_CHECK_EXPR(SUP_IOCTL_CALL_VMMR0_BIG, pVMMReq->u32Magic == SUPVMMR0REQHDR_MAGIC);
1399	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_VMMR0_BIG, SUP_IOCTL_CALL_VMMR0_BIG_SIZE_IN(pVMMReq->cbReq), SUP_IOCTL_CALL_VMMR0_BIG_SIZE_OUT(pVMMReq->cbReq));
1400
1401	/* execute */
1402	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
1403	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pReq->u.In.pVMR0, pReq->u.In.idCpu, pReq->u.In.uOperation, pVMMReq, pReq->u.In.u64Arg, pSession);
1404	else
1405	pReq->Hdr.rc = VERR_WRONG_ORDER;
1406
1407	if ( RT_FAILURE(pReq->Hdr.rc)
1408	&& pReq->Hdr.rc != VERR_INTERRUPTED
1409	&& pReq->Hdr.rc != VERR_TIMEOUT)
1410	Log(("SUP_IOCTL_CALL_VMMR0_BIG: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1411	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1412	else
1413	Log4(("SUP_IOCTL_CALL_VMMR0_BIG: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1414	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1415	return 0;
1416	}
1417
1418	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GET_PAGING_MODE):
1419	{
1420	/* validate */
1421	PSUPGETPAGINGMODE pReq = (PSUPGETPAGINGMODE)pReqHdr;
1422	REQ_CHECK_SIZES(SUP_IOCTL_GET_PAGING_MODE);
1423
1424	/* execute */
1425	pReq->Hdr.rc = VINF_SUCCESS;
1426	pReq->u.Out.enmMode = SUPR0GetPagingMode();
1427	return 0;
1428	}
1429
1430	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOW_ALLOC):
1431	{
1432	/* validate */
1433	PSUPLOWALLOC pReq = (PSUPLOWALLOC)pReqHdr;
1434	REQ_CHECK_EXPR(SUP_IOCTL_LOW_ALLOC, pReq->Hdr.cbIn <= SUP_IOCTL_LOW_ALLOC_SIZE_IN);
1435	REQ_CHECK_SIZES_EX(SUP_IOCTL_LOW_ALLOC, SUP_IOCTL_LOW_ALLOC_SIZE_IN, SUP_IOCTL_LOW_ALLOC_SIZE_OUT(pReq->u.In.cPages));
1436
1437	/* execute */
1438	pReq->Hdr.rc = SUPR0LowAlloc(pSession, pReq->u.In.cPages, &pReq->u.Out.pvR0, &pReq->u.Out.pvR3, &pReq->u.Out.aPages[0]);
1439	if (RT_FAILURE(pReq->Hdr.rc))
1440	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1441	return 0;
1442	}
1443
1444	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOW_FREE):
1445	{
1446	/* validate */
1447	PSUPLOWFREE pReq = (PSUPLOWFREE)pReqHdr;
1448	REQ_CHECK_SIZES(SUP_IOCTL_LOW_FREE);
1449
1450	/* execute */
1451	pReq->Hdr.rc = SUPR0LowFree(pSession, (RTHCUINTPTR)pReq->u.In.pvR3);
1452	return 0;
1453	}
1454
1455	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GIP_MAP):
1456	{
1457	/* validate */
1458	PSUPGIPMAP pReq = (PSUPGIPMAP)pReqHdr;
1459	REQ_CHECK_SIZES(SUP_IOCTL_GIP_MAP);
1460
1461	/* execute */
1462	pReq->Hdr.rc = SUPR0GipMap(pSession, &pReq->u.Out.pGipR3, &pReq->u.Out.HCPhysGip);
1463	if (RT_SUCCESS(pReq->Hdr.rc))
1464	pReq->u.Out.pGipR0 = pDevExt->pGip;
1465	return 0;
1466	}
1467
1468	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GIP_UNMAP):
1469	{
1470	/* validate */
1471	PSUPGIPUNMAP pReq = (PSUPGIPUNMAP)pReqHdr;
1472	REQ_CHECK_SIZES(SUP_IOCTL_GIP_UNMAP);
1473
1474	/* execute */
1475	pReq->Hdr.rc = SUPR0GipUnmap(pSession);
1476	return 0;
1477	}
1478
1479	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_SET_VM_FOR_FAST):
1480	{
1481	/* validate */
1482	PSUPSETVMFORFAST pReq = (PSUPSETVMFORFAST)pReqHdr;
1483	REQ_CHECK_SIZES(SUP_IOCTL_SET_VM_FOR_FAST);
1484	REQ_CHECK_EXPR_FMT( !pReq->u.In.pVMR0
1485	\|\| ( VALID_PTR(pReq->u.In.pVMR0)
1486	&& !((uintptr_t)pReq->u.In.pVMR0 & (PAGE_SIZE - 1))),
1487	("SUP_IOCTL_SET_VM_FOR_FAST: pVMR0=%p!\n", pReq->u.In.pVMR0));
1488	/* execute */
1489	pSession->pVM = pReq->u.In.pVMR0;
1490	pReq->Hdr.rc = VINF_SUCCESS;
1491	return 0;
1492	}
1493
1494	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_ALLOC_EX):
1495	{
1496	/* validate */
1497	PSUPPAGEALLOCEX pReq = (PSUPPAGEALLOCEX)pReqHdr;
1498	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_ALLOC_EX, pReq->Hdr.cbIn <= SUP_IOCTL_PAGE_ALLOC_EX_SIZE_IN);
1499	REQ_CHECK_SIZES_EX(SUP_IOCTL_PAGE_ALLOC_EX, SUP_IOCTL_PAGE_ALLOC_EX_SIZE_IN, SUP_IOCTL_PAGE_ALLOC_EX_SIZE_OUT(pReq->u.In.cPages));
1500	REQ_CHECK_EXPR_FMT(pReq->u.In.fKernelMapping \|\| pReq->u.In.fUserMapping,
1501	("SUP_IOCTL_PAGE_ALLOC_EX: No mapping requested!\n"));
1502	REQ_CHECK_EXPR_FMT(pReq->u.In.fUserMapping,
1503	("SUP_IOCTL_PAGE_ALLOC_EX: Must have user mapping!\n"));
1504	REQ_CHECK_EXPR_FMT(!pReq->u.In.fReserved0 && !pReq->u.In.fReserved1,
1505	("SUP_IOCTL_PAGE_ALLOC_EX: fReserved0=%d fReserved1=%d\n", pReq->u.In.fReserved0, pReq->u.In.fReserved1));
1506
1507	/* execute */
1508	pReq->Hdr.rc = SUPR0PageAllocEx(pSession, pReq->u.In.cPages, 0 /* fFlags */,
1509	pReq->u.In.fUserMapping ? &pReq->u.Out.pvR3 : NULL,
1510	pReq->u.In.fKernelMapping ? &pReq->u.Out.pvR0 : NULL,
1511	&pReq->u.Out.aPages[0]);
1512	if (RT_FAILURE(pReq->Hdr.rc))
1513	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1514	return 0;
1515	}
1516
1517	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_MAP_KERNEL):
1518	{
1519	/* validate */
1520	PSUPPAGEMAPKERNEL pReq = (PSUPPAGEMAPKERNEL)pReqHdr;
1521	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_MAP_KERNEL);
1522	REQ_CHECK_EXPR_FMT(!pReq->u.In.fFlags, ("SUP_IOCTL_PAGE_MAP_KERNEL: fFlags=%#x! MBZ\n", pReq->u.In.fFlags));
1523	REQ_CHECK_EXPR_FMT(!(pReq->u.In.offSub & PAGE_OFFSET_MASK), ("SUP_IOCTL_PAGE_MAP_KERNEL: offSub=%#x\n", pReq->u.In.offSub));
1524	REQ_CHECK_EXPR_FMT(pReq->u.In.cbSub && !(pReq->u.In.cbSub & PAGE_OFFSET_MASK),
1525	("SUP_IOCTL_PAGE_MAP_KERNEL: cbSub=%#x\n", pReq->u.In.cbSub));
1526
1527	/* execute */
1528	pReq->Hdr.rc = SUPR0PageMapKernel(pSession, pReq->u.In.pvR3, pReq->u.In.offSub, pReq->u.In.cbSub,
1529	pReq->u.In.fFlags, &pReq->u.Out.pvR0);
1530	if (RT_FAILURE(pReq->Hdr.rc))
1531	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1532	return 0;
1533	}
1534
1535	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_PROTECT):
1536	{
1537	/* validate */
1538	PSUPPAGEPROTECT pReq = (PSUPPAGEPROTECT)pReqHdr;
1539	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_PROTECT);
1540	REQ_CHECK_EXPR_FMT(!(pReq->u.In.fProt & ~(RTMEM_PROT_READ \| RTMEM_PROT_WRITE \| RTMEM_PROT_EXEC \| RTMEM_PROT_NONE)),
1541	("SUP_IOCTL_PAGE_PROTECT: fProt=%#x!\n", pReq->u.In.fProt));
1542	REQ_CHECK_EXPR_FMT(!(pReq->u.In.offSub & PAGE_OFFSET_MASK), ("SUP_IOCTL_PAGE_PROTECT: offSub=%#x\n", pReq->u.In.offSub));
1543	REQ_CHECK_EXPR_FMT(pReq->u.In.cbSub && !(pReq->u.In.cbSub & PAGE_OFFSET_MASK),
1544	("SUP_IOCTL_PAGE_PROTECT: cbSub=%#x\n", pReq->u.In.cbSub));
1545
1546	/* execute */
1547	pReq->Hdr.rc = SUPR0PageProtect(pSession, pReq->u.In.pvR3, pReq->u.In.pvR0, pReq->u.In.offSub, pReq->u.In.cbSub, pReq->u.In.fProt);
1548	return 0;
1549	}
1550
1551	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_FREE):
1552	{
1553	/* validate */
1554	PSUPPAGEFREE pReq = (PSUPPAGEFREE)pReqHdr;
1555	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_FREE);
1556
1557	/* execute */
1558	pReq->Hdr.rc = SUPR0PageFree(pSession, pReq->u.In.pvR3);
1559	return 0;
1560	}
1561
1562	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CALL_SERVICE(0)):
1563	{
1564	/* validate */
1565	PSUPCALLSERVICE pReq = (PSUPCALLSERVICE)pReqHdr;
1566	Log4(("SUP_IOCTL_CALL_SERVICE: op=%u in=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1567	pReq->u.In.uOperation, pReq->Hdr.cbIn, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1568
1569	if (pReq->Hdr.cbIn == SUP_IOCTL_CALL_SERVICE_SIZE(0))
1570	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_SERVICE, SUP_IOCTL_CALL_SERVICE_SIZE_IN(0), SUP_IOCTL_CALL_SERVICE_SIZE_OUT(0));
1571	else
1572	{
1573	PSUPR0SERVICEREQHDR pSrvReq = (PSUPR0SERVICEREQHDR)&pReq->abReqPkt[0];
1574	REQ_CHECK_EXPR_FMT(pReq->Hdr.cbIn >= SUP_IOCTL_CALL_SERVICE_SIZE(sizeof(SUPR0SERVICEREQHDR)),
1575	("SUP_IOCTL_CALL_SERVICE: cbIn=%#x < %#lx\n", pReq->Hdr.cbIn, SUP_IOCTL_CALL_SERVICE_SIZE(sizeof(SUPR0SERVICEREQHDR))));
1576	REQ_CHECK_EXPR(SUP_IOCTL_CALL_SERVICE, pSrvReq->u32Magic == SUPR0SERVICEREQHDR_MAGIC);
1577	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_SERVICE, SUP_IOCTL_CALL_SERVICE_SIZE_IN(pSrvReq->cbReq), SUP_IOCTL_CALL_SERVICE_SIZE_OUT(pSrvReq->cbReq));
1578	}
1579	REQ_CHECK_EXPR(SUP_IOCTL_CALL_SERVICE, RTStrEnd(pReq->u.In.szName, sizeof(pReq->u.In.szName)));
1580
1581	/* execute */
1582	pReq->Hdr.rc = supdrvIOCtl_CallServiceModule(pDevExt, pSession, pReq);
1583	return 0;
1584	}
1585
1586	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOGGER_SETTINGS(0)):
1587	{
1588	/* validate */
1589	PSUPLOGGERSETTINGS pReq = (PSUPLOGGERSETTINGS)pReqHdr;
1590	size_t cbStrTab;
1591	REQ_CHECK_SIZE_OUT(SUP_IOCTL_LOGGER_SETTINGS, SUP_IOCTL_LOGGER_SETTINGS_SIZE_OUT);
1592	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->Hdr.cbIn >= SUP_IOCTL_LOGGER_SETTINGS_SIZE_IN(1));
1593	cbStrTab = pReq->Hdr.cbIn - SUP_IOCTL_LOGGER_SETTINGS_SIZE_IN(0);
1594	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.offGroups < cbStrTab);
1595	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.offFlags < cbStrTab);
1596	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.offDestination < cbStrTab);
1597	REQ_CHECK_EXPR_FMT(pReq->u.In.szStrings[cbStrTab - 1] == '\0',
1598	("SUP_IOCTL_LOGGER_SETTINGS: cbIn=%#x cbStrTab=%#zx LastChar=%d\n",
1599	pReq->Hdr.cbIn, cbStrTab, pReq->u.In.szStrings[cbStrTab - 1]));
1600	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.fWhich <= SUPLOGGERSETTINGS_WHICH_RELEASE);
1601	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.fWhat <= SUPLOGGERSETTINGS_WHAT_DESTROY);
1602
1603	/* execute */
1604	pReq->Hdr.rc = supdrvIOCtl_LoggerSettings(pDevExt, pSession, pReq);
1605	return 0;
1606	}
1607
1608	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_SEM_OP2):
1609	{
1610	/* validate */
1611	PSUPSEMOP2 pReq = (PSUPSEMOP2)pReqHdr;
1612	REQ_CHECK_SIZES_EX(SUP_IOCTL_SEM_OP2, SUP_IOCTL_SEM_OP2_SIZE_IN, SUP_IOCTL_SEM_OP2_SIZE_OUT);
1613	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP2, pReq->u.In.uReserved == 0);
1614
1615	/* execute */
1616	switch (pReq->u.In.uType)
1617	{
1618	case SUP_SEM_TYPE_EVENT:
1619	{
1620	SUPSEMEVENT hEvent = (SUPSEMEVENT)(uintptr_t)pReq->u.In.hSem;
1621	switch (pReq->u.In.uOp)
1622	{
1623	case SUPSEMOP2_WAIT_MS_REL:
1624	pReq->Hdr.rc = SUPSemEventWaitNoResume(pSession, hEvent, pReq->u.In.uArg.cRelMsTimeout);
1625	break;
1626	case SUPSEMOP2_WAIT_NS_ABS:
1627	pReq->Hdr.rc = SUPSemEventWaitNsAbsIntr(pSession, hEvent, pReq->u.In.uArg.uAbsNsTimeout);
1628	break;
1629	case SUPSEMOP2_WAIT_NS_REL:
1630	pReq->Hdr.rc = SUPSemEventWaitNsRelIntr(pSession, hEvent, pReq->u.In.uArg.cRelNsTimeout);
1631	break;
1632	case SUPSEMOP2_SIGNAL:
1633	pReq->Hdr.rc = SUPSemEventSignal(pSession, hEvent);
1634	break;
1635	case SUPSEMOP2_CLOSE:
1636	pReq->Hdr.rc = SUPSemEventClose(pSession, hEvent);
1637	break;
1638	case SUPSEMOP2_RESET:
1639	default:
1640	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1641	break;
1642	}
1643	break;
1644	}
1645
1646	case SUP_SEM_TYPE_EVENT_MULTI:
1647	{
1648	SUPSEMEVENTMULTI hEventMulti = (SUPSEMEVENTMULTI)(uintptr_t)pReq->u.In.hSem;
1649	switch (pReq->u.In.uOp)
1650	{
1651	case SUPSEMOP2_WAIT_MS_REL:
1652	pReq->Hdr.rc = SUPSemEventMultiWaitNoResume(pSession, hEventMulti, pReq->u.In.uArg.cRelMsTimeout);
1653	break;
1654	case SUPSEMOP2_WAIT_NS_ABS:
1655	pReq->Hdr.rc = SUPSemEventMultiWaitNsAbsIntr(pSession, hEventMulti, pReq->u.In.uArg.uAbsNsTimeout);
1656	break;
1657	case SUPSEMOP2_WAIT_NS_REL:
1658	pReq->Hdr.rc = SUPSemEventMultiWaitNsRelIntr(pSession, hEventMulti, pReq->u.In.uArg.cRelNsTimeout);
1659	break;
1660	case SUPSEMOP2_SIGNAL:
1661	pReq->Hdr.rc = SUPSemEventMultiSignal(pSession, hEventMulti);
1662	break;
1663	case SUPSEMOP2_CLOSE:
1664	pReq->Hdr.rc = SUPSemEventMultiClose(pSession, hEventMulti);
1665	break;
1666	case SUPSEMOP2_RESET:
1667	pReq->Hdr.rc = SUPSemEventMultiReset(pSession, hEventMulti);
1668	break;
1669	default:
1670	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1671	break;
1672	}
1673	break;
1674	}
1675
1676	default:
1677	pReq->Hdr.rc = VERR_INVALID_PARAMETER;
1678	break;
1679	}
1680	return 0;
1681	}
1682
1683	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_SEM_OP3):
1684	{
1685	/* validate */
1686	PSUPSEMOP3 pReq = (PSUPSEMOP3)pReqHdr;
1687	REQ_CHECK_SIZES_EX(SUP_IOCTL_SEM_OP3, SUP_IOCTL_SEM_OP3_SIZE_IN, SUP_IOCTL_SEM_OP3_SIZE_OUT);
1688	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, pReq->u.In.u32Reserved == 0 && pReq->u.In.u64Reserved == 0);
1689
1690	/* execute */
1691	switch (pReq->u.In.uType)
1692	{
1693	case SUP_SEM_TYPE_EVENT:
1694	{
1695	SUPSEMEVENT hEvent = (SUPSEMEVENT)(uintptr_t)pReq->u.In.hSem;
1696	switch (pReq->u.In.uOp)
1697	{
1698	case SUPSEMOP3_CREATE:
1699	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEvent == NIL_SUPSEMEVENT);
1700	pReq->Hdr.rc = SUPSemEventCreate(pSession, &hEvent);
1701	pReq->u.Out.hSem = (uint32_t)(uintptr_t)hEvent;
1702	break;
1703	case SUPSEMOP3_GET_RESOLUTION:
1704	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEvent == NIL_SUPSEMEVENT);
1705	pReq->Hdr.rc = VINF_SUCCESS;
1706	pReq->Hdr.cbOut = sizeof(*pReq);
1707	pReq->u.Out.cNsResolution = SUPSemEventGetResolution(pSession);
1708	break;
1709	default:
1710	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1711	break;
1712	}
1713	break;
1714	}
1715
1716	case SUP_SEM_TYPE_EVENT_MULTI:
1717	{
1718	SUPSEMEVENTMULTI hEventMulti = (SUPSEMEVENTMULTI)(uintptr_t)pReq->u.In.hSem;
1719	switch (pReq->u.In.uOp)
1720	{
1721	case SUPSEMOP3_CREATE:
1722	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEventMulti == NIL_SUPSEMEVENTMULTI);
1723	pReq->Hdr.rc = SUPSemEventMultiCreate(pSession, &hEventMulti);
1724	pReq->u.Out.hSem = (uint32_t)(uintptr_t)hEventMulti;
1725	break;
1726	case SUPSEMOP3_GET_RESOLUTION:
1727	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEventMulti == NIL_SUPSEMEVENTMULTI);
1728	pReq->Hdr.rc = VINF_SUCCESS;
1729	pReq->u.Out.cNsResolution = SUPSemEventMultiGetResolution(pSession);
1730	break;
1731	default:
1732	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1733	break;
1734	}
1735	break;
1736	}
1737
1738	default:
1739	pReq->Hdr.rc = VERR_INVALID_PARAMETER;
1740	break;
1741	}
1742	return 0;
1743	}
1744
1745	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_VT_CAPS):
1746	{
1747	/* validate */
1748	PSUPVTCAPS pReq = (PSUPVTCAPS)pReqHdr;
1749	REQ_CHECK_SIZES(SUP_IOCTL_VT_CAPS);
1750	REQ_CHECK_EXPR(SUP_IOCTL_VT_CAPS, pReq->Hdr.cbIn <= SUP_IOCTL_VT_CAPS_SIZE_IN);
1751
1752	/* execute */
1753	pReq->Hdr.rc = SUPR0QueryVTCaps(pSession, &pReq->u.Out.Caps);
1754	if (RT_FAILURE(pReq->Hdr.rc))
1755	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1756	return 0;
1757	}
1758
1759	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_TRACER_OPEN):
1760	{
1761	/* validate */
1762	PSUPTRACEROPEN pReq = (PSUPTRACEROPEN)pReqHdr;
1763	REQ_CHECK_SIZES(SUP_IOCTL_TRACER_OPEN);
1764
1765	/* execute */
1766	pReq->Hdr.rc = supdrvIOCtl_TracerOpen(pDevExt, pSession, pReq->u.In.uCookie, pReq->u.In.uArg);
1767	return 0;
1768	}
1769
1770	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_TRACER_CLOSE):
1771	{
1772	/* validate */
1773	REQ_CHECK_SIZES(SUP_IOCTL_TRACER_CLOSE);
1774
1775	/* execute */
1776	pReqHdr->rc = supdrvIOCtl_TracerClose(pDevExt, pSession);
1777	return 0;
1778	}
1779
1780	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_TRACER_IOCTL):
1781	{
1782	/* validate */
1783	PSUPTRACERIOCTL pReq = (PSUPTRACERIOCTL)pReqHdr;
1784	REQ_CHECK_SIZES(SUP_IOCTL_TRACER_IOCTL);
1785
1786	/* execute */
1787	pReqHdr->rc = supdrvIOCtl_TracerIOCtl(pDevExt, pSession, pReq->u.In.uCmd, pReq->u.In.uArg, &pReq->u.Out.iRetVal);
1788	return 0;
1789	}
1790
1791	default:
1792	Log(("Unknown IOCTL %#lx\n", (long)uIOCtl));
1793	break;
1794	}
1795	return VERR_GENERAL_FAILURE;
1796	}
1797
1798
1799	/**
1800	* I/O Control worker.
1801	*
1802	* @returns IPRT status code.
1803	* @retval VERR_INVALID_PARAMETER if the request is invalid.
1804	*
1805	* @param uIOCtl Function number.
1806	* @param pDevExt Device extention.
1807	* @param pSession Session data.
1808	* @param pReqHdr The request header.
1809	*/
1810	int VBOXCALL supdrvIOCtl(uintptr_t uIOCtl, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPREQHDR pReqHdr)
1811	{
1812	int rc;
1813	VBOXDRV_IOCTL_ENTRY(pSession, uIOCtl, pReqHdr);
1814
1815	/*
1816	* Validate the request.
1817	*/
1818	/* this first check could probably be omitted as its also done by the OS specific code... */
1819	if (RT_UNLIKELY( (pReqHdr->fFlags & SUPREQHDR_FLAGS_MAGIC_MASK) != SUPREQHDR_FLAGS_MAGIC
1820	\|\| pReqHdr->cbIn < sizeof(*pReqHdr)
1821	\|\| pReqHdr->cbOut < sizeof(*pReqHdr)))
1822	{
1823	OSDBGPRINT(("vboxdrv: Bad ioctl request header; cbIn=%#lx cbOut=%#lx fFlags=%#lx\n",
1824	(long)pReqHdr->cbIn, (long)pReqHdr->cbOut, (long)pReqHdr->fFlags));
1825	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VINF_SUCCESS);
1826	return VERR_INVALID_PARAMETER;
1827	}
1828	if (RT_UNLIKELY(!RT_VALID_PTR(pSession)))
1829	{
1830	OSDBGPRINT(("vboxdrv: Invalid pSession valud %p (ioctl=%p)\n", pSession, (void *)uIOCtl));
1831	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VINF_SUCCESS);
1832	return VERR_INVALID_PARAMETER;
1833	}
1834	if (RT_UNLIKELY(uIOCtl == SUP_IOCTL_COOKIE))
1835	{
1836	if (pReqHdr->u32Cookie != SUPCOOKIE_INITIAL_COOKIE)
1837	{
1838	OSDBGPRINT(("SUP_IOCTL_COOKIE: bad cookie %#lx\n", (long)pReqHdr->u32Cookie));
1839	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VINF_SUCCESS);
1840	return VERR_INVALID_PARAMETER;
1841	}
1842	}
1843	else if (RT_UNLIKELY( pReqHdr->u32Cookie != pDevExt->u32Cookie
1844	\|\| pReqHdr->u32SessionCookie != pSession->u32Cookie))
1845	{
1846	OSDBGPRINT(("vboxdrv: bad cookie %#lx / %#lx.\n", (long)pReqHdr->u32Cookie, (long)pReqHdr->u32SessionCookie));
1847	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VINF_SUCCESS);
1848	return VERR_INVALID_PARAMETER;
1849	}
1850
1851	/*
1852	* Hand it to an inner function to avoid lots of unnecessary return tracepoints
1853	*/
1854	rc = supdrvIOCtlInner(uIOCtl, pDevExt, pSession, pReqHdr);
1855
1856	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, pReqHdr->rc, rc);
1857	return rc;
1858	}
1859
1860
1861	/**
1862	* Inter-Driver Communication (IDC) worker.
1863	*
1864	* @returns VBox status code.
1865	* @retval VINF_SUCCESS on success.
1866	* @retval VERR_INVALID_PARAMETER if the request is invalid.
1867	* @retval VERR_NOT_SUPPORTED if the request isn't supported.
1868	*
1869	* @param uReq The request (function) code.
1870	* @param pDevExt Device extention.
1871	* @param pSession Session data.
1872	* @param pReqHdr The request header.
1873	*/
1874	int VBOXCALL supdrvIDC(uintptr_t uReq, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVIDCREQHDR pReqHdr)
1875	{
1876	/*
1877	* The OS specific code has already validated the pSession
1878	* pointer, and the request size being greater or equal to
1879	* size of the header.
1880	*
1881	* So, just check that pSession is a kernel context session.
1882	*/
1883	if (RT_UNLIKELY( pSession
1884	&& pSession->R0Process != NIL_RTR0PROCESS))
1885	return VERR_INVALID_PARAMETER;
1886
1887	/*
1888	* Validation macro.
1889	*/
1890	#define REQ_CHECK_IDC_SIZE(Name, cbExpect) \
1891	do { \
1892	if (RT_UNLIKELY(pReqHdr->cb != (cbExpect))) \
1893	{ \
1894	OSDBGPRINT(( #Name ": Invalid input/output sizes. cb=%ld expected %ld.\n", \
1895	(long)pReqHdr->cb, (long)(cbExpect))); \
1896	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
1897	} \
1898	} while (0)
1899
1900	switch (uReq)
1901	{
1902	case SUPDRV_IDC_REQ_CONNECT:
1903	{
1904	PSUPDRVIDCREQCONNECT pReq = (PSUPDRVIDCREQCONNECT)pReqHdr;
1905	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_CONNECT, sizeof(*pReq));
1906
1907	/*
1908	* Validate the cookie and other input.
1909	*/
1910	if (pReq->Hdr.pSession != NULL)
1911	{
1912	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: Hdr.pSession=%p expected NULL!\n", pReq->Hdr.pSession));
1913	return pReqHdr->rc = VERR_INVALID_PARAMETER;
1914	}
1915	if (pReq->u.In.u32MagicCookie != SUPDRVIDCREQ_CONNECT_MAGIC_COOKIE)
1916	{
1917	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: u32MagicCookie=%#x expected %#x!\n",
1918	(unsigned)pReq->u.In.u32MagicCookie, (unsigned)SUPDRVIDCREQ_CONNECT_MAGIC_COOKIE));
1919	return pReqHdr->rc = VERR_INVALID_PARAMETER;
1920	}
1921	if ( pReq->u.In.uMinVersion > pReq->u.In.uReqVersion
1922	\|\| (pReq->u.In.uMinVersion & UINT32_C(0xffff0000)) != (pReq->u.In.uReqVersion & UINT32_C(0xffff0000)))
1923	{
1924	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: uMinVersion=%#x uMaxVersion=%#x doesn't match!\n",
1925	pReq->u.In.uMinVersion, pReq->u.In.uReqVersion));
1926	return pReqHdr->rc = VERR_INVALID_PARAMETER;
1927	}
1928	if (pSession != NULL)
1929	{
1930	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: pSession=%p expected NULL!\n", pSession));
1931	return pReqHdr->rc = VERR_INVALID_PARAMETER;
1932	}
1933
1934	/*
1935	* Match the version.
1936	* The current logic is very simple, match the major interface version.
1937	*/
1938	if ( pReq->u.In.uMinVersion > SUPDRV_IDC_VERSION
1939	\|\| (pReq->u.In.uMinVersion & 0xffff0000) != (SUPDRV_IDC_VERSION & 0xffff0000))
1940	{
1941	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: Version mismatch. Requested: %#x Min: %#x Current: %#x\n",
1942	pReq->u.In.uReqVersion, pReq->u.In.uMinVersion, (unsigned)SUPDRV_IDC_VERSION));
1943	pReq->u.Out.pSession = NULL;
1944	pReq->u.Out.uSessionVersion = 0xffffffff;
1945	pReq->u.Out.uDriverVersion = SUPDRV_IDC_VERSION;
1946	pReq->u.Out.uDriverRevision = VBOX_SVN_REV;
1947	pReq->Hdr.rc = VERR_VERSION_MISMATCH;
1948	return VINF_SUCCESS;
1949	}
1950
1951	pReq->u.Out.pSession = NULL;
1952	pReq->u.Out.uSessionVersion = SUPDRV_IDC_VERSION;
1953	pReq->u.Out.uDriverVersion = SUPDRV_IDC_VERSION;
1954	pReq->u.Out.uDriverRevision = VBOX_SVN_REV;
1955
1956	pReq->Hdr.rc = supdrvCreateSession(pDevExt, false /* fUser */, &pSession);
1957	if (RT_FAILURE(pReq->Hdr.rc))
1958	{
1959	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: failed to create session, rc=%d\n", pReq->Hdr.rc));
1960	return VINF_SUCCESS;
1961	}
1962
1963	pReq->u.Out.pSession = pSession;
1964	pReq->Hdr.pSession = pSession;
1965
1966	return VINF_SUCCESS;
1967	}
1968
1969	case SUPDRV_IDC_REQ_DISCONNECT:
1970	{
1971	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_DISCONNECT, sizeof(*pReqHdr));
1972
1973	supdrvCloseSession(pDevExt, pSession);
1974	return pReqHdr->rc = VINF_SUCCESS;
1975	}
1976
1977	case SUPDRV_IDC_REQ_GET_SYMBOL:
1978	{
1979	PSUPDRVIDCREQGETSYM pReq = (PSUPDRVIDCREQGETSYM)pReqHdr;
1980	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_GET_SYMBOL, sizeof(*pReq));
1981
1982	pReq->Hdr.rc = supdrvIDC_LdrGetSymbol(pDevExt, pSession, pReq);
1983	return VINF_SUCCESS;
1984	}
1985
1986	case SUPDRV_IDC_REQ_COMPONENT_REGISTER_FACTORY:
1987	{
1988	PSUPDRVIDCREQCOMPREGFACTORY pReq = (PSUPDRVIDCREQCOMPREGFACTORY)pReqHdr;
1989	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_COMPONENT_REGISTER_FACTORY, sizeof(*pReq));
1990
1991	pReq->Hdr.rc = SUPR0ComponentRegisterFactory(pSession, pReq->u.In.pFactory);
1992	return VINF_SUCCESS;
1993	}
1994
1995	case SUPDRV_IDC_REQ_COMPONENT_DEREGISTER_FACTORY:
1996	{
1997	PSUPDRVIDCREQCOMPDEREGFACTORY pReq = (PSUPDRVIDCREQCOMPDEREGFACTORY)pReqHdr;
1998	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_COMPONENT_DEREGISTER_FACTORY, sizeof(*pReq));
1999
2000	pReq->Hdr.rc = SUPR0ComponentDeregisterFactory(pSession, pReq->u.In.pFactory);
2001	return VINF_SUCCESS;
2002	}
2003
2004	default:
2005	Log(("Unknown IDC %#lx\n", (long)uReq));
2006	break;
2007	}
2008
2009	#undef REQ_CHECK_IDC_SIZE
2010	return VERR_NOT_SUPPORTED;
2011	}
2012
2013
2014	/**
2015	* Register a object for reference counting.
2016	* The object is registered with one reference in the specified session.
2017	*
2018	* @returns Unique identifier on success (pointer).
2019	* All future reference must use this identifier.
2020	* @returns NULL on failure.
2021	* @param pfnDestructor The destructore function which will be called when the reference count reaches 0.
2022	* @param pvUser1 The first user argument.
2023	* @param pvUser2 The second user argument.
2024	*/
2025	SUPR0DECL(void ) SUPR0ObjRegister(PSUPDRVSESSION pSession, SUPDRVOBJTYPE enmType, PFNSUPDRVDESTRUCTOR pfnDestructor, void pvUser1, void *pvUser2)
2026	{
2027	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
2028	PSUPDRVOBJ pObj;
2029	PSUPDRVUSAGE pUsage;
2030
2031	/*
2032	* Validate the input.
2033	*/
2034	AssertReturn(SUP_IS_SESSION_VALID(pSession), NULL);
2035	AssertReturn(enmType > SUPDRVOBJTYPE_INVALID && enmType < SUPDRVOBJTYPE_END, NULL);
2036	AssertPtrReturn(pfnDestructor, NULL);
2037
2038	/*
2039	* Allocate and initialize the object.
2040	*/
2041	pObj = (PSUPDRVOBJ)RTMemAlloc(sizeof(*pObj));
2042	if (!pObj)
2043	return NULL;
2044	pObj->u32Magic = SUPDRVOBJ_MAGIC;
2045	pObj->enmType = enmType;
2046	pObj->pNext = NULL;
2047	pObj->cUsage = 1;
2048	pObj->pfnDestructor = pfnDestructor;
2049	pObj->pvUser1 = pvUser1;
2050	pObj->pvUser2 = pvUser2;
2051	pObj->CreatorUid = pSession->Uid;
2052	pObj->CreatorGid = pSession->Gid;
2053	pObj->CreatorProcess= pSession->Process;
2054	supdrvOSObjInitCreator(pObj, pSession);
2055
2056	/*
2057	* Allocate the usage record.
2058	* (We keep freed usage records around to simplify SUPR0ObjAddRefEx().)
2059	*/
2060	RTSpinlockAcquire(pDevExt->Spinlock);
2061
2062	pUsage = pDevExt->pUsageFree;
2063	if (pUsage)
2064	pDevExt->pUsageFree = pUsage->pNext;
2065	else
2066	{
2067	RTSpinlockRelease(pDevExt->Spinlock);
2068	pUsage = (PSUPDRVUSAGE)RTMemAlloc(sizeof(*pUsage));
2069	if (!pUsage)
2070	{
2071	RTMemFree(pObj);
2072	return NULL;
2073	}
2074	RTSpinlockAcquire(pDevExt->Spinlock);
2075	}
2076
2077	/*
2078	* Insert the object and create the session usage record.
2079	*/
2080	/* The object. */
2081	pObj->pNext = pDevExt->pObjs;
2082	pDevExt->pObjs = pObj;
2083
2084	/* The session record. */
2085	pUsage->cUsage = 1;
2086	pUsage->pObj = pObj;
2087	pUsage->pNext = pSession->pUsage;
2088	/* Log2(("SUPR0ObjRegister: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext)); */
2089	pSession->pUsage = pUsage;
2090
2091	RTSpinlockRelease(pDevExt->Spinlock);
2092
2093	Log(("SUPR0ObjRegister: returns %p (pvUser1=%p, pvUser=%p)\n", pObj, pvUser1, pvUser2));
2094	return pObj;
2095	}
2096
2097
2098	/**
2099	* Increment the reference counter for the object associating the reference
2100	* with the specified session.
2101	*
2102	* @returns IPRT status code.
2103	* @param pvObj The identifier returned by SUPR0ObjRegister().
2104	* @param pSession The session which is referencing the object.
2105	*
2106	* @remarks The caller should not own any spinlocks and must carefully protect
2107	* itself against potential race with the destructor so freed memory
2108	* isn't accessed here.
2109	*/
2110	SUPR0DECL(int) SUPR0ObjAddRef(void *pvObj, PSUPDRVSESSION pSession)
2111	{
2112	return SUPR0ObjAddRefEx(pvObj, pSession, false /* fNoBlocking */);
2113	}
2114
2115
2116	/**
2117	* Increment the reference counter for the object associating the reference
2118	* with the specified session.
2119	*
2120	* @returns IPRT status code.
2121	* @retval VERR_TRY_AGAIN if fNoBlocking was set and a new usage record
2122	* couldn't be allocated. (If you see this you're not doing the right
2123	* thing and it won't ever work reliably.)
2124	*
2125	* @param pvObj The identifier returned by SUPR0ObjRegister().
2126	* @param pSession The session which is referencing the object.
2127	* @param fNoBlocking Set if it's not OK to block. Never try to make the
2128	* first reference to an object in a session with this
2129	* argument set.
2130	*
2131	* @remarks The caller should not own any spinlocks and must carefully protect
2132	* itself against potential race with the destructor so freed memory
2133	* isn't accessed here.
2134	*/
2135	SUPR0DECL(int) SUPR0ObjAddRefEx(void *pvObj, PSUPDRVSESSION pSession, bool fNoBlocking)
2136	{
2137	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
2138	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
2139	int rc = VINF_SUCCESS;
2140	PSUPDRVUSAGE pUsagePre;
2141	PSUPDRVUSAGE pUsage;
2142
2143	/*
2144	* Validate the input.
2145	* Be ready for the destruction race (someone might be stuck in the
2146	* destructor waiting a lock we own).
2147	*/
2148	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2149	AssertPtrReturn(pObj, VERR_INVALID_POINTER);
2150	AssertMsgReturn(pObj->u32Magic == SUPDRVOBJ_MAGIC \|\| pObj->u32Magic == SUPDRVOBJ_MAGIC_DEAD,
2151	("Invalid pvObj=%p magic=%#x (expected %#x or %#x)\n", pvObj, pObj->u32Magic, SUPDRVOBJ_MAGIC, SUPDRVOBJ_MAGIC_DEAD),
2152	VERR_INVALID_PARAMETER);
2153
2154	RTSpinlockAcquire(pDevExt->Spinlock);
2155
2156	if (RT_UNLIKELY(pObj->u32Magic != SUPDRVOBJ_MAGIC))
2157	{
2158	RTSpinlockRelease(pDevExt->Spinlock);
2159
2160	AssertMsgFailed(("pvObj=%p magic=%#x\n", pvObj, pObj->u32Magic));
2161	return VERR_WRONG_ORDER;
2162	}
2163
2164	/*
2165	* Preallocate the usage record if we can.
2166	*/
2167	pUsagePre = pDevExt->pUsageFree;
2168	if (pUsagePre)
2169	pDevExt->pUsageFree = pUsagePre->pNext;
2170	else if (!fNoBlocking)
2171	{
2172	RTSpinlockRelease(pDevExt->Spinlock);
2173	pUsagePre = (PSUPDRVUSAGE)RTMemAlloc(sizeof(*pUsagePre));
2174	if (!pUsagePre)
2175	return VERR_NO_MEMORY;
2176
2177	RTSpinlockAcquire(pDevExt->Spinlock);
2178	if (RT_UNLIKELY(pObj->u32Magic != SUPDRVOBJ_MAGIC))
2179	{
2180	RTSpinlockRelease(pDevExt->Spinlock);
2181
2182	AssertMsgFailed(("pvObj=%p magic=%#x\n", pvObj, pObj->u32Magic));
2183	return VERR_WRONG_ORDER;
2184	}
2185	}
2186
2187	/*
2188	* Reference the object.
2189	*/
2190	pObj->cUsage++;
2191
2192	/*
2193	* Look for the session record.
2194	*/
2195	for (pUsage = pSession->pUsage; pUsage; pUsage = pUsage->pNext)
2196	{
2197	/Log(("SUPR0AddRef: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext));/
2198	if (pUsage->pObj == pObj)
2199	break;
2200	}
2201	if (pUsage)
2202	pUsage->cUsage++;
2203	else if (pUsagePre)
2204	{
2205	/* create a new session record. */
2206	pUsagePre->cUsage = 1;
2207	pUsagePre->pObj = pObj;
2208	pUsagePre->pNext = pSession->pUsage;
2209	pSession->pUsage = pUsagePre;
2210	/Log(("SUPR0AddRef: pUsagePre=%p:{.pObj=%p, .pNext=%p}\n", pUsagePre, pUsagePre->pObj, pUsagePre->pNext));/
2211
2212	pUsagePre = NULL;
2213	}
2214	else
2215	{
2216	pObj->cUsage--;
2217	rc = VERR_TRY_AGAIN;
2218	}
2219
2220	/*
2221	* Put any unused usage record into the free list..
2222	*/
2223	if (pUsagePre)
2224	{
2225	pUsagePre->pNext = pDevExt->pUsageFree;
2226	pDevExt->pUsageFree = pUsagePre;
2227	}
2228
2229	RTSpinlockRelease(pDevExt->Spinlock);
2230
2231	return rc;
2232	}
2233
2234
2235	/**
2236	* Decrement / destroy a reference counter record for an object.
2237	*
2238	* The object is uniquely identified by pfnDestructor+pvUser1+pvUser2.
2239	*
2240	* @returns IPRT status code.
2241	* @retval VINF_SUCCESS if not destroyed.
2242	* @retval VINF_OBJECT_DESTROYED if it's destroyed by this release call.
2243	* @retval VERR_INVALID_PARAMETER if the object isn't valid. Will assert in
2244	* string builds.
2245	*
2246	* @param pvObj The identifier returned by SUPR0ObjRegister().
2247	* @param pSession The session which is referencing the object.
2248	*/
2249	SUPR0DECL(int) SUPR0ObjRelease(void *pvObj, PSUPDRVSESSION pSession)
2250	{
2251	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
2252	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
2253	int rc = VERR_INVALID_PARAMETER;
2254	PSUPDRVUSAGE pUsage;
2255	PSUPDRVUSAGE pUsagePrev;
2256
2257	/*
2258	* Validate the input.
2259	*/
2260	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2261	AssertMsgReturn(VALID_PTR(pObj) && pObj->u32Magic == SUPDRVOBJ_MAGIC,
2262	("Invalid pvObj=%p magic=%#x (exepcted %#x)\n", pvObj, pObj ? pObj->u32Magic : 0, SUPDRVOBJ_MAGIC),
2263	VERR_INVALID_PARAMETER);
2264
2265	/*
2266	* Acquire the spinlock and look for the usage record.
2267	*/
2268	RTSpinlockAcquire(pDevExt->Spinlock);
2269
2270	for (pUsagePrev = NULL, pUsage = pSession->pUsage;
2271	pUsage;
2272	pUsagePrev = pUsage, pUsage = pUsage->pNext)
2273	{
2274	/Log2(("SUPR0ObjRelease: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext));/
2275	if (pUsage->pObj == pObj)
2276	{
2277	rc = VINF_SUCCESS;
2278	AssertMsg(pUsage->cUsage >= 1 && pObj->cUsage >= pUsage->cUsage, ("glob %d; sess %d\n", pObj->cUsage, pUsage->cUsage));
2279	if (pUsage->cUsage > 1)
2280	{
2281	pObj->cUsage--;
2282	pUsage->cUsage--;
2283	}
2284	else
2285	{
2286	/*
2287	* Free the session record.
2288	*/
2289	if (pUsagePrev)
2290	pUsagePrev->pNext = pUsage->pNext;
2291	else
2292	pSession->pUsage = pUsage->pNext;
2293	pUsage->pNext = pDevExt->pUsageFree;
2294	pDevExt->pUsageFree = pUsage;
2295
2296	/* What about the object? */
2297	if (pObj->cUsage > 1)
2298	pObj->cUsage--;
2299	else
2300	{
2301	/*
2302	* Object is to be destroyed, unlink it.
2303	*/
2304	pObj->u32Magic = SUPDRVOBJ_MAGIC_DEAD;
2305	rc = VINF_OBJECT_DESTROYED;
2306	if (pDevExt->pObjs == pObj)
2307	pDevExt->pObjs = pObj->pNext;
2308	else
2309	{
2310	PSUPDRVOBJ pObjPrev;
2311	for (pObjPrev = pDevExt->pObjs; pObjPrev; pObjPrev = pObjPrev->pNext)
2312	if (pObjPrev->pNext == pObj)
2313	{
2314	pObjPrev->pNext = pObj->pNext;
2315	break;
2316	}
2317	Assert(pObjPrev);
2318	}
2319	}
2320	}
2321	break;
2322	}
2323	}
2324
2325	RTSpinlockRelease(pDevExt->Spinlock);
2326
2327	/*
2328	* Call the destructor and free the object if required.
2329	*/
2330	if (rc == VINF_OBJECT_DESTROYED)
2331	{
2332	Log(("SUPR0ObjRelease: destroying %p/%d (%p/%p) cpid=%RTproc pid=%RTproc dtor=%p\n",
2333	pObj, pObj->enmType, pObj->pvUser1, pObj->pvUser2, pObj->CreatorProcess, RTProcSelf(), pObj->pfnDestructor));
2334	if (pObj->pfnDestructor)
2335	pObj->pfnDestructor(pObj, pObj->pvUser1, pObj->pvUser2);
2336	RTMemFree(pObj);
2337	}
2338
2339	AssertMsg(pUsage, ("pvObj=%p\n", pvObj));
2340	return rc;
2341	}
2342
2343
2344	/**
2345	* Verifies that the current process can access the specified object.
2346	*
2347	* @returns The following IPRT status code:
2348	* @retval VINF_SUCCESS if access was granted.
2349	* @retval VERR_PERMISSION_DENIED if denied access.
2350	* @retval VERR_INVALID_PARAMETER if invalid parameter.
2351	*
2352	* @param pvObj The identifier returned by SUPR0ObjRegister().
2353	* @param pSession The session which wishes to access the object.
2354	* @param pszObjName Object string name. This is optional and depends on the object type.
2355	*
2356	* @remark The caller is responsible for making sure the object isn't removed while
2357	* we're inside this function. If uncertain about this, just call AddRef before calling us.
2358	*/
2359	SUPR0DECL(int) SUPR0ObjVerifyAccess(void pvObj, PSUPDRVSESSION pSession, const char pszObjName)
2360	{
2361	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
2362	int rc;
2363
2364	/*
2365	* Validate the input.
2366	*/
2367	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2368	AssertMsgReturn(VALID_PTR(pObj) && pObj->u32Magic == SUPDRVOBJ_MAGIC,
2369	("Invalid pvObj=%p magic=%#x (exepcted %#x)\n", pvObj, pObj ? pObj->u32Magic : 0, SUPDRVOBJ_MAGIC),
2370	VERR_INVALID_PARAMETER);
2371
2372	/*
2373	* Check access. (returns true if a decision has been made.)
2374	*/
2375	rc = VERR_INTERNAL_ERROR;
2376	if (supdrvOSObjCanAccess(pObj, pSession, pszObjName, &rc))
2377	return rc;
2378
2379	/*
2380	* Default policy is to allow the user to access his own
2381	* stuff but nothing else.
2382	*/
2383	if (pObj->CreatorUid == pSession->Uid)
2384	return VINF_SUCCESS;
2385	return VERR_PERMISSION_DENIED;
2386	}
2387
2388
2389	/**
2390	* Lock pages.
2391	*
2392	* @returns IPRT status code.
2393	* @param pSession Session to which the locked memory should be associated.
2394	* @param pvR3 Start of the memory range to lock.
2395	* This must be page aligned.
2396	* @param cPages Number of pages to lock.
2397	* @param paPages Where to put the physical addresses of locked memory.
2398	*/
2399	SUPR0DECL(int) SUPR0LockMem(PSUPDRVSESSION pSession, RTR3PTR pvR3, uint32_t cPages, PRTHCPHYS paPages)
2400	{
2401	int rc;
2402	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2403	const size_t cb = (size_t)cPages << PAGE_SHIFT;
2404	LogFlow(("SUPR0LockMem: pSession=%p pvR3=%p cPages=%d paPages=%p\n", pSession, (void *)pvR3, cPages, paPages));
2405
2406	/*
2407	* Verify input.
2408	*/
2409	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2410	AssertPtrReturn(paPages, VERR_INVALID_PARAMETER);
2411	if ( RT_ALIGN_R3PT(pvR3, PAGE_SIZE, RTR3PTR) != pvR3
2412	\|\| !pvR3)
2413	{
2414	Log(("pvR3 (%p) must be page aligned and not NULL!\n", (void *)pvR3));
2415	return VERR_INVALID_PARAMETER;
2416	}
2417
2418	/*
2419	* Let IPRT do the job.
2420	*/
2421	Mem.eType = MEMREF_TYPE_LOCKED;
2422	rc = RTR0MemObjLockUser(&Mem.MemObj, pvR3, cb, RTMEM_PROT_READ \| RTMEM_PROT_WRITE, RTR0ProcHandleSelf());
2423	if (RT_SUCCESS(rc))
2424	{
2425	uint32_t iPage = cPages;
2426	AssertMsg(RTR0MemObjAddressR3(Mem.MemObj) == pvR3, ("%p == %p\n", RTR0MemObjAddressR3(Mem.MemObj), pvR3));
2427	AssertMsg(RTR0MemObjSize(Mem.MemObj) == cb, ("%x == %x\n", RTR0MemObjSize(Mem.MemObj), cb));
2428
2429	while (iPage-- > 0)
2430	{
2431	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MemObj, iPage);
2432	if (RT_UNLIKELY(paPages[iPage] == NIL_RTCCPHYS))
2433	{
2434	AssertMsgFailed(("iPage=%d\n", iPage));
2435	rc = VERR_INTERNAL_ERROR;
2436	break;
2437	}
2438	}
2439	if (RT_SUCCESS(rc))
2440	rc = supdrvMemAdd(&Mem, pSession);
2441	if (RT_FAILURE(rc))
2442	{
2443	int rc2 = RTR0MemObjFree(Mem.MemObj, false);
2444	AssertRC(rc2);
2445	}
2446	}
2447
2448	return rc;
2449	}
2450
2451
2452	/**
2453	* Unlocks the memory pointed to by pv.
2454	*
2455	* @returns IPRT status code.
2456	* @param pSession Session to which the memory was locked.
2457	* @param pvR3 Memory to unlock.
2458	*/
2459	SUPR0DECL(int) SUPR0UnlockMem(PSUPDRVSESSION pSession, RTR3PTR pvR3)
2460	{
2461	LogFlow(("SUPR0UnlockMem: pSession=%p pvR3=%p\n", pSession, (void *)pvR3));
2462	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2463	return supdrvMemRelease(pSession, (RTHCUINTPTR)pvR3, MEMREF_TYPE_LOCKED);
2464	}
2465
2466
2467	/**
2468	* Allocates a chunk of page aligned memory with contiguous and fixed physical
2469	* backing.
2470	*
2471	* @returns IPRT status code.
2472	* @param pSession Session data.
2473	* @param cPages Number of pages to allocate.
2474	* @param ppvR0 Where to put the address of Ring-0 mapping the allocated memory.
2475	* @param ppvR3 Where to put the address of Ring-3 mapping the allocated memory.
2476	* @param pHCPhys Where to put the physical address of allocated memory.
2477	*/
2478	SUPR0DECL(int) SUPR0ContAlloc(PSUPDRVSESSION pSession, uint32_t cPages, PRTR0PTR ppvR0, PRTR3PTR ppvR3, PRTHCPHYS pHCPhys)
2479	{
2480	int rc;
2481	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2482	LogFlow(("SUPR0ContAlloc: pSession=%p cPages=%d ppvR0=%p ppvR3=%p pHCPhys=%p\n", pSession, cPages, ppvR0, ppvR3, pHCPhys));
2483
2484	/*
2485	* Validate input.
2486	*/
2487	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2488	if (!ppvR3 \|\| !ppvR0 \|\| !pHCPhys)
2489	{
2490	Log(("Null pointer. All of these should be set: pSession=%p ppvR0=%p ppvR3=%p pHCPhys=%p\n",
2491	pSession, ppvR0, ppvR3, pHCPhys));
2492	return VERR_INVALID_PARAMETER;
2493
2494	}
2495	if (cPages < 1 \|\| cPages >= 256)
2496	{
2497	Log(("Illegal request cPages=%d, must be greater than 0 and smaller than 256.\n", cPages));
2498	return VERR_PAGE_COUNT_OUT_OF_RANGE;
2499	}
2500
2501	/*
2502	* Let IPRT do the job.
2503	*/
2504	rc = RTR0MemObjAllocCont(&Mem.MemObj, cPages << PAGE_SHIFT, true /* executable R0 mapping */);
2505	if (RT_SUCCESS(rc))
2506	{
2507	int rc2;
2508	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2509	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2510	if (RT_SUCCESS(rc))
2511	{
2512	Mem.eType = MEMREF_TYPE_CONT;
2513	rc = supdrvMemAdd(&Mem, pSession);
2514	if (!rc)
2515	{
2516	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2517	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2518	*pHCPhys = RTR0MemObjGetPagePhysAddr(Mem.MemObj, 0);
2519	return 0;
2520	}
2521
2522	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2523	AssertRC(rc2);
2524	}
2525	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2526	AssertRC(rc2);
2527	}
2528
2529	return rc;
2530	}
2531
2532
2533	/**
2534	* Frees memory allocated using SUPR0ContAlloc().
2535	*
2536	* @returns IPRT status code.
2537	* @param pSession The session to which the memory was allocated.
2538	* @param uPtr Pointer to the memory (ring-3 or ring-0).
2539	*/
2540	SUPR0DECL(int) SUPR0ContFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
2541	{
2542	LogFlow(("SUPR0ContFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
2543	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2544	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_CONT);
2545	}
2546
2547
2548	/**
2549	* Allocates a chunk of page aligned memory with fixed physical backing below 4GB.
2550	*
2551	* The memory isn't zeroed.
2552	*
2553	* @returns IPRT status code.
2554	* @param pSession Session data.
2555	* @param cPages Number of pages to allocate.
2556	* @param ppvR0 Where to put the address of Ring-0 mapping of the allocated memory.
2557	* @param ppvR3 Where to put the address of Ring-3 mapping of the allocated memory.
2558	* @param paPages Where to put the physical addresses of allocated memory.
2559	*/
2560	SUPR0DECL(int) SUPR0LowAlloc(PSUPDRVSESSION pSession, uint32_t cPages, PRTR0PTR ppvR0, PRTR3PTR ppvR3, PRTHCPHYS paPages)
2561	{
2562	unsigned iPage;
2563	int rc;
2564	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2565	LogFlow(("SUPR0LowAlloc: pSession=%p cPages=%d ppvR3=%p ppvR0=%p paPages=%p\n", pSession, cPages, ppvR3, ppvR0, paPages));
2566
2567	/*
2568	* Validate input.
2569	*/
2570	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2571	if (!ppvR3 \|\| !ppvR0 \|\| !paPages)
2572	{
2573	Log(("Null pointer. All of these should be set: pSession=%p ppvR3=%p ppvR0=%p paPages=%p\n",
2574	pSession, ppvR3, ppvR0, paPages));
2575	return VERR_INVALID_PARAMETER;
2576
2577	}
2578	if (cPages < 1 \|\| cPages >= 256)
2579	{
2580	Log(("Illegal request cPages=%d, must be greater than 0 and smaller than 256.\n", cPages));
2581	return VERR_PAGE_COUNT_OUT_OF_RANGE;
2582	}
2583
2584	/*
2585	* Let IPRT do the work.
2586	*/
2587	rc = RTR0MemObjAllocLow(&Mem.MemObj, cPages << PAGE_SHIFT, true /* executable ring-0 mapping */);
2588	if (RT_SUCCESS(rc))
2589	{
2590	int rc2;
2591	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2592	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2593	if (RT_SUCCESS(rc))
2594	{
2595	Mem.eType = MEMREF_TYPE_LOW;
2596	rc = supdrvMemAdd(&Mem, pSession);
2597	if (!rc)
2598	{
2599	for (iPage = 0; iPage < cPages; iPage++)
2600	{
2601	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MemObj, iPage);
2602	AssertMsg(!(paPages[iPage] & (PAGE_SIZE - 1)), ("iPage=%d Phys=%RHp\n", paPages[iPage]));
2603	}
2604	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2605	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2606	return 0;
2607	}
2608
2609	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2610	AssertRC(rc2);
2611	}
2612
2613	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2614	AssertRC(rc2);
2615	}
2616
2617	return rc;
2618	}
2619
2620
2621	/**
2622	* Frees memory allocated using SUPR0LowAlloc().
2623	*
2624	* @returns IPRT status code.
2625	* @param pSession The session to which the memory was allocated.
2626	* @param uPtr Pointer to the memory (ring-3 or ring-0).
2627	*/
2628	SUPR0DECL(int) SUPR0LowFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
2629	{
2630	LogFlow(("SUPR0LowFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
2631	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2632	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_LOW);
2633	}
2634
2635
2636
2637	/**
2638	* Allocates a chunk of memory with both R0 and R3 mappings.
2639	* The memory is fixed and it's possible to query the physical addresses using SUPR0MemGetPhys().
2640	*
2641	* @returns IPRT status code.
2642	* @param pSession The session to associated the allocation with.
2643	* @param cb Number of bytes to allocate.
2644	* @param ppvR0 Where to store the address of the Ring-0 mapping.
2645	* @param ppvR3 Where to store the address of the Ring-3 mapping.
2646	*/
2647	SUPR0DECL(int) SUPR0MemAlloc(PSUPDRVSESSION pSession, uint32_t cb, PRTR0PTR ppvR0, PRTR3PTR ppvR3)
2648	{
2649	int rc;
2650	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2651	LogFlow(("SUPR0MemAlloc: pSession=%p cb=%d ppvR0=%p ppvR3=%p\n", pSession, cb, ppvR0, ppvR3));
2652
2653	/*
2654	* Validate input.
2655	*/
2656	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2657	AssertPtrReturn(ppvR0, VERR_INVALID_POINTER);
2658	AssertPtrReturn(ppvR3, VERR_INVALID_POINTER);
2659	if (cb < 1 \|\| cb >= _4M)
2660	{
2661	Log(("Illegal request cb=%u; must be greater than 0 and smaller than 4MB.\n", cb));
2662	return VERR_INVALID_PARAMETER;
2663	}
2664
2665	/*
2666	* Let IPRT do the work.
2667	*/
2668	rc = RTR0MemObjAllocPage(&Mem.MemObj, cb, true /* executable ring-0 mapping */);
2669	if (RT_SUCCESS(rc))
2670	{
2671	int rc2;
2672	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2673	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2674	if (RT_SUCCESS(rc))
2675	{
2676	Mem.eType = MEMREF_TYPE_MEM;
2677	rc = supdrvMemAdd(&Mem, pSession);
2678	if (!rc)
2679	{
2680	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2681	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2682	return VINF_SUCCESS;
2683	}
2684
2685	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2686	AssertRC(rc2);
2687	}
2688
2689	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2690	AssertRC(rc2);
2691	}
2692
2693	return rc;
2694	}
2695
2696
2697	/**
2698	* Get the physical addresses of memory allocated using SUPR0MemAlloc().
2699	*
2700	* @returns IPRT status code.
2701	* @param pSession The session to which the memory was allocated.
2702	* @param uPtr The Ring-0 or Ring-3 address returned by SUPR0MemAlloc().
2703	* @param paPages Where to store the physical addresses.
2704	*/
2705	SUPR0DECL(int) SUPR0MemGetPhys(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, PSUPPAGE paPages) /** @todo switch this bugger to RTHCPHYS */
2706	{
2707	PSUPDRVBUNDLE pBundle;
2708	LogFlow(("SUPR0MemGetPhys: pSession=%p uPtr=%p paPages=%p\n", pSession, (void *)uPtr, paPages));
2709
2710	/*
2711	* Validate input.
2712	*/
2713	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2714	AssertPtrReturn(paPages, VERR_INVALID_POINTER);
2715	AssertReturn(uPtr, VERR_INVALID_PARAMETER);
2716
2717	/*
2718	* Search for the address.
2719	*/
2720	RTSpinlockAcquire(pSession->Spinlock);
2721	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
2722	{
2723	if (pBundle->cUsed > 0)
2724	{
2725	unsigned i;
2726	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
2727	{
2728	if ( pBundle->aMem[i].eType == MEMREF_TYPE_MEM
2729	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2730	&& ( (RTHCUINTPTR)RTR0MemObjAddress(pBundle->aMem[i].MemObj) == uPtr
2731	\|\| ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
2732	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == uPtr)
2733	)
2734	)
2735	{
2736	const size_t cPages = RTR0MemObjSize(pBundle->aMem[i].MemObj) >> PAGE_SHIFT;
2737	size_t iPage;
2738	for (iPage = 0; iPage < cPages; iPage++)
2739	{
2740	paPages[iPage].Phys = RTR0MemObjGetPagePhysAddr(pBundle->aMem[i].MemObj, iPage);
2741	paPages[iPage].uReserved = 0;
2742	}
2743	RTSpinlockRelease(pSession->Spinlock);
2744	return VINF_SUCCESS;
2745	}
2746	}
2747	}
2748	}
2749	RTSpinlockRelease(pSession->Spinlock);
2750	Log(("Failed to find %p!!!\n", (void *)uPtr));
2751	return VERR_INVALID_PARAMETER;
2752	}
2753
2754
2755	/**
2756	* Free memory allocated by SUPR0MemAlloc().
2757	*
2758	* @returns IPRT status code.
2759	* @param pSession The session owning the allocation.
2760	* @param uPtr The Ring-0 or Ring-3 address returned by SUPR0MemAlloc().
2761	*/
2762	SUPR0DECL(int) SUPR0MemFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
2763	{
2764	LogFlow(("SUPR0MemFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
2765	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2766	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_MEM);
2767	}
2768
2769
2770	/**
2771	* Allocates a chunk of memory with a kernel or/and a user mode mapping.
2772	*
2773	* The memory is fixed and it's possible to query the physical addresses using
2774	* SUPR0MemGetPhys().
2775	*
2776	* @returns IPRT status code.
2777	* @param pSession The session to associated the allocation with.
2778	* @param cPages The number of pages to allocate.
2779	* @param fFlags Flags, reserved for the future. Must be zero.
2780	* @param ppvR3 Where to store the address of the Ring-3 mapping.
2781	* NULL if no ring-3 mapping.
2782	* @param ppvR3 Where to store the address of the Ring-0 mapping.
2783	* NULL if no ring-0 mapping.
2784	* @param paPages Where to store the addresses of the pages. Optional.
2785	*/
2786	SUPR0DECL(int) SUPR0PageAllocEx(PSUPDRVSESSION pSession, uint32_t cPages, uint32_t fFlags, PRTR3PTR ppvR3, PRTR0PTR ppvR0, PRTHCPHYS paPages)
2787	{
2788	int rc;
2789	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2790	LogFlow(("SUPR0PageAlloc: pSession=%p cb=%d ppvR3=%p\n", pSession, cPages, ppvR3));
2791
2792	/*
2793	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
2794	*/
2795	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2796	AssertPtrNullReturn(ppvR3, VERR_INVALID_POINTER);
2797	AssertPtrNullReturn(ppvR0, VERR_INVALID_POINTER);
2798	AssertReturn(ppvR3 \|\| ppvR0, VERR_INVALID_PARAMETER);
2799	AssertReturn(!fFlags, VERR_INVALID_PARAMETER);
2800	if (cPages < 1 \|\| cPages > VBOX_MAX_ALLOC_PAGE_COUNT)
2801	{
2802	Log(("SUPR0PageAlloc: Illegal request cb=%u; must be greater than 0 and smaller than %uMB (VBOX_MAX_ALLOC_PAGE_COUNT pages).\n", cPages, VBOX_MAX_ALLOC_PAGE_COUNT * (_1M / _4K)));
2803	return VERR_PAGE_COUNT_OUT_OF_RANGE;
2804	}
2805
2806	/*
2807	* Let IPRT do the work.
2808	*/
2809	if (ppvR0)
2810	rc = RTR0MemObjAllocPage(&Mem.MemObj, (size_t)cPages * PAGE_SIZE, true /* fExecutable */);
2811	else
2812	rc = RTR0MemObjAllocPhysNC(&Mem.MemObj, (size_t)cPages * PAGE_SIZE, NIL_RTHCPHYS);
2813	if (RT_SUCCESS(rc))
2814	{
2815	int rc2;
2816	if (ppvR3)
2817	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2818	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2819	else
2820	Mem.MapObjR3 = NIL_RTR0MEMOBJ;
2821	if (RT_SUCCESS(rc))
2822	{
2823	Mem.eType = MEMREF_TYPE_PAGE;
2824	rc = supdrvMemAdd(&Mem, pSession);
2825	if (!rc)
2826	{
2827	if (ppvR3)
2828	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2829	if (ppvR0)
2830	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2831	if (paPages)
2832	{
2833	uint32_t iPage = cPages;
2834	while (iPage-- > 0)
2835	{
2836	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MapObjR3, iPage);
2837	Assert(paPages[iPage] != NIL_RTHCPHYS);
2838	}
2839	}
2840	return VINF_SUCCESS;
2841	}
2842
2843	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2844	AssertRC(rc2);
2845	}
2846
2847	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2848	AssertRC(rc2);
2849	}
2850	return rc;
2851	}
2852
2853
2854	/**
2855	* Maps a chunk of memory previously allocated by SUPR0PageAllocEx into kernel
2856	* space.
2857	*
2858	* @returns IPRT status code.
2859	* @param pSession The session to associated the allocation with.
2860	* @param pvR3 The ring-3 address returned by SUPR0PageAllocEx.
2861	* @param offSub Where to start mapping. Must be page aligned.
2862	* @param cbSub How much to map. Must be page aligned.
2863	* @param fFlags Flags, MBZ.
2864	* @param ppvR0 Where to return the address of the ring-0 mapping on
2865	* success.
2866	*/
2867	SUPR0DECL(int) SUPR0PageMapKernel(PSUPDRVSESSION pSession, RTR3PTR pvR3, uint32_t offSub, uint32_t cbSub,
2868	uint32_t fFlags, PRTR0PTR ppvR0)
2869	{
2870	int rc;
2871	PSUPDRVBUNDLE pBundle;
2872	RTR0MEMOBJ hMemObj = NIL_RTR0MEMOBJ;
2873	LogFlow(("SUPR0PageMapKernel: pSession=%p pvR3=%p offSub=%#x cbSub=%#x\n", pSession, pvR3, offSub, cbSub));
2874
2875	/*
2876	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
2877	*/
2878	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2879	AssertPtrNullReturn(ppvR0, VERR_INVALID_POINTER);
2880	AssertReturn(!fFlags, VERR_INVALID_PARAMETER);
2881	AssertReturn(!(offSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
2882	AssertReturn(!(cbSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
2883	AssertReturn(cbSub, VERR_INVALID_PARAMETER);
2884
2885	/*
2886	* Find the memory object.
2887	*/
2888	RTSpinlockAcquire(pSession->Spinlock);
2889	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
2890	{
2891	if (pBundle->cUsed > 0)
2892	{
2893	unsigned i;
2894	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
2895	{
2896	if ( ( pBundle->aMem[i].eType == MEMREF_TYPE_PAGE
2897	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2898	&& pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
2899	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == pvR3)
2900	\|\| ( pBundle->aMem[i].eType == MEMREF_TYPE_LOCKED
2901	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2902	&& pBundle->aMem[i].MapObjR3 == NIL_RTR0MEMOBJ
2903	&& RTR0MemObjAddressR3(pBundle->aMem[i].MemObj) == pvR3))
2904	{
2905	hMemObj = pBundle->aMem[i].MemObj;
2906	break;
2907	}
2908	}
2909	}
2910	}
2911	RTSpinlockRelease(pSession->Spinlock);
2912
2913	rc = VERR_INVALID_PARAMETER;
2914	if (hMemObj != NIL_RTR0MEMOBJ)
2915	{
2916	/*
2917	* Do some further input validations before calling IPRT.
2918	* (Cleanup is done indirectly by telling RTR0MemObjFree to include mappings.)
2919	*/
2920	size_t cbMemObj = RTR0MemObjSize(hMemObj);
2921	if ( offSub < cbMemObj
2922	&& cbSub <= cbMemObj
2923	&& offSub + cbSub <= cbMemObj)
2924	{
2925	RTR0MEMOBJ hMapObj;
2926	rc = RTR0MemObjMapKernelEx(&hMapObj, hMemObj, (void *)-1, 0,
2927	RTMEM_PROT_READ \| RTMEM_PROT_WRITE, offSub, cbSub);
2928	if (RT_SUCCESS(rc))
2929	*ppvR0 = RTR0MemObjAddress(hMapObj);
2930	}
2931	else
2932	SUPR0Printf("SUPR0PageMapKernel: cbMemObj=%#x offSub=%#x cbSub=%#x\n", cbMemObj, offSub, cbSub);
2933
2934	}
2935	return rc;
2936	}
2937
2938
2939	/**
2940	* Changes the page level protection of one or more pages previously allocated
2941	* by SUPR0PageAllocEx.
2942	*
2943	* @returns IPRT status code.
2944	* @param pSession The session to associated the allocation with.
2945	* @param pvR3 The ring-3 address returned by SUPR0PageAllocEx.
2946	* NIL_RTR3PTR if the ring-3 mapping should be unaffected.
2947	* @param pvR0 The ring-0 address returned by SUPR0PageAllocEx.
2948	* NIL_RTR0PTR if the ring-0 mapping should be unaffected.
2949	* @param offSub Where to start changing. Must be page aligned.
2950	* @param cbSub How much to change. Must be page aligned.
2951	* @param fProt The new page level protection, see RTMEM_PROT_*.
2952	*/
2953	SUPR0DECL(int) SUPR0PageProtect(PSUPDRVSESSION pSession, RTR3PTR pvR3, RTR0PTR pvR0, uint32_t offSub, uint32_t cbSub, uint32_t fProt)
2954	{
2955	int rc;
2956	PSUPDRVBUNDLE pBundle;
2957	RTR0MEMOBJ hMemObjR0 = NIL_RTR0MEMOBJ;
2958	RTR0MEMOBJ hMemObjR3 = NIL_RTR0MEMOBJ;
2959	LogFlow(("SUPR0PageProtect: pSession=%p pvR3=%p pvR0=%p offSub=%#x cbSub=%#x fProt-%#x\n", pSession, pvR3, pvR0, offSub, cbSub, fProt));
2960
2961	/*
2962	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
2963	*/
2964	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2965	AssertReturn(!(fProt & ~(RTMEM_PROT_READ \| RTMEM_PROT_WRITE \| RTMEM_PROT_EXEC \| RTMEM_PROT_NONE)), VERR_INVALID_PARAMETER);
2966	AssertReturn(!(offSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
2967	AssertReturn(!(cbSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
2968	AssertReturn(cbSub, VERR_INVALID_PARAMETER);
2969
2970	/*
2971	* Find the memory object.
2972	*/
2973	RTSpinlockAcquire(pSession->Spinlock);
2974	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
2975	{
2976	if (pBundle->cUsed > 0)
2977	{
2978	unsigned i;
2979	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
2980	{
2981	if ( pBundle->aMem[i].eType == MEMREF_TYPE_PAGE
2982	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2983	&& ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
2984	\|\| pvR3 == NIL_RTR3PTR)
2985	&& ( pvR0 == NIL_RTR0PTR
2986	\|\| RTR0MemObjAddress(pBundle->aMem[i].MemObj) == pvR0)
2987	&& ( pvR3 == NIL_RTR3PTR
2988	\|\| RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == pvR3))
2989	{
2990	if (pvR0 != NIL_RTR0PTR)
2991	hMemObjR0 = pBundle->aMem[i].MemObj;
2992	if (pvR3 != NIL_RTR3PTR)
2993	hMemObjR3 = pBundle->aMem[i].MapObjR3;
2994	break;
2995	}
2996	}
2997	}
2998	}
2999	RTSpinlockRelease(pSession->Spinlock);
3000
3001	rc = VERR_INVALID_PARAMETER;
3002	if ( hMemObjR0 != NIL_RTR0MEMOBJ
3003	\|\| hMemObjR3 != NIL_RTR0MEMOBJ)
3004	{
3005	/*
3006	* Do some further input validations before calling IPRT.
3007	*/
3008	size_t cbMemObj = hMemObjR0 != NIL_RTR0PTR ? RTR0MemObjSize(hMemObjR0) : RTR0MemObjSize(hMemObjR3);
3009	if ( offSub < cbMemObj
3010	&& cbSub <= cbMemObj
3011	&& offSub + cbSub <= cbMemObj)
3012	{
3013	rc = VINF_SUCCESS;
3014	if (hMemObjR3 != NIL_RTR0PTR)
3015	rc = RTR0MemObjProtect(hMemObjR3, offSub, cbSub, fProt);
3016	if (hMemObjR0 != NIL_RTR0PTR && RT_SUCCESS(rc))
3017	rc = RTR0MemObjProtect(hMemObjR0, offSub, cbSub, fProt);
3018	}
3019	else
3020	SUPR0Printf("SUPR0PageMapKernel: cbMemObj=%#x offSub=%#x cbSub=%#x\n", cbMemObj, offSub, cbSub);
3021
3022	}
3023	return rc;
3024
3025	}
3026
3027
3028	/**
3029	* Free memory allocated by SUPR0PageAlloc() and SUPR0PageAllocEx().
3030	*
3031	* @returns IPRT status code.
3032	* @param pSession The session owning the allocation.
3033	* @param pvR3 The Ring-3 address returned by SUPR0PageAlloc() or
3034	* SUPR0PageAllocEx().
3035	*/
3036	SUPR0DECL(int) SUPR0PageFree(PSUPDRVSESSION pSession, RTR3PTR pvR3)
3037	{
3038	LogFlow(("SUPR0PageFree: pSession=%p pvR3=%p\n", pSession, (void *)pvR3));
3039	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3040	return supdrvMemRelease(pSession, (RTHCUINTPTR)pvR3, MEMREF_TYPE_PAGE);
3041	}
3042
3043
3044	/**
3045	* Gets the paging mode of the current CPU.
3046	*
3047	* @returns Paging mode, SUPPAGEINGMODE_INVALID on error.
3048	*/
3049	SUPR0DECL(SUPPAGINGMODE) SUPR0GetPagingMode(void)
3050	{
3051	SUPPAGINGMODE enmMode;
3052
3053	RTR0UINTREG cr0 = ASMGetCR0();
3054	if ((cr0 & (X86_CR0_PG \| X86_CR0_PE)) != (X86_CR0_PG \| X86_CR0_PE))
3055	enmMode = SUPPAGINGMODE_INVALID;
3056	else
3057	{
3058	RTR0UINTREG cr4 = ASMGetCR4();
3059	uint32_t fNXEPlusLMA = 0;
3060	if (cr4 & X86_CR4_PAE)
3061	{
3062	uint32_t fAmdFeatures = ASMCpuId_EDX(0x80000001);
3063	if (fAmdFeatures & (X86_CPUID_AMD_FEATURE_EDX_NX \| X86_CPUID_AMD_FEATURE_EDX_LONG_MODE))
3064	{
3065	uint64_t efer = ASMRdMsr(MSR_K6_EFER);
3066	if ((fAmdFeatures & X86_CPUID_AMD_FEATURE_EDX_NX) && (efer & MSR_K6_EFER_NXE))
3067	fNXEPlusLMA \|= RT_BIT(0);
3068	if ((fAmdFeatures & X86_CPUID_AMD_FEATURE_EDX_LONG_MODE) && (efer & MSR_K6_EFER_LMA))
3069	fNXEPlusLMA \|= RT_BIT(1);
3070	}
3071	}
3072
3073	switch ((cr4 & (X86_CR4_PAE \| X86_CR4_PGE)) \| fNXEPlusLMA)
3074	{
3075	case 0:
3076	enmMode = SUPPAGINGMODE_32_BIT;
3077	break;
3078
3079	case X86_CR4_PGE:
3080	enmMode = SUPPAGINGMODE_32_BIT_GLOBAL;
3081	break;
3082
3083	case X86_CR4_PAE:
3084	enmMode = SUPPAGINGMODE_PAE;
3085	break;
3086
3087	case X86_CR4_PAE \| RT_BIT(0):
3088	enmMode = SUPPAGINGMODE_PAE_NX;
3089	break;
3090
3091	case X86_CR4_PAE \| X86_CR4_PGE:
3092	enmMode = SUPPAGINGMODE_PAE_GLOBAL;
3093	break;
3094
3095	case X86_CR4_PAE \| X86_CR4_PGE \| RT_BIT(0):
3096	enmMode = SUPPAGINGMODE_PAE_GLOBAL;
3097	break;
3098
3099	case RT_BIT(1) \| X86_CR4_PAE:
3100	enmMode = SUPPAGINGMODE_AMD64;
3101	break;
3102
3103	case RT_BIT(1) \| X86_CR4_PAE \| RT_BIT(0):
3104	enmMode = SUPPAGINGMODE_AMD64_NX;
3105	break;
3106
3107	case RT_BIT(1) \| X86_CR4_PAE \| X86_CR4_PGE:
3108	enmMode = SUPPAGINGMODE_AMD64_GLOBAL;
3109	break;
3110
3111	case RT_BIT(1) \| X86_CR4_PAE \| X86_CR4_PGE \| RT_BIT(0):
3112	enmMode = SUPPAGINGMODE_AMD64_GLOBAL_NX;
3113	break;
3114
3115	default:
3116	AssertMsgFailed(("Cannot happen! cr4=%#x fNXEPlusLMA=%d\n", cr4, fNXEPlusLMA));
3117	enmMode = SUPPAGINGMODE_INVALID;
3118	break;
3119	}
3120	}
3121	return enmMode;
3122	}
3123
3124
3125	/**
3126	* Enables or disabled hardware virtualization extensions using native OS APIs.
3127	*
3128	* @returns VBox status code.
3129	* @retval VINF_SUCCESS on success.
3130	* @retval VERR_NOT_SUPPORTED if not supported by the native OS.
3131	*
3132	* @param fEnable Whether to enable or disable.
3133	*/
3134	SUPR0DECL(int) SUPR0EnableVTx(bool fEnable)
3135	{
3136	#ifdef RT_OS_DARWIN
3137	return supdrvOSEnableVTx(fEnable);
3138	#else
3139	return VERR_NOT_SUPPORTED;
3140	#endif
3141	}
3142
3143
3144	/** @todo document me */
3145	SUPR0DECL(int) SUPR0QueryVTCaps(PSUPDRVSESSION pSession, uint32_t *pfCaps)
3146	{
3147	/*
3148	* Input validation.
3149	*/
3150	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3151	AssertPtrReturn(pfCaps, VERR_INVALID_POINTER);
3152
3153	*pfCaps = 0;
3154
3155	if (ASMHasCpuId())
3156	{
3157	uint32_t u32FeaturesECX;
3158	uint32_t u32Dummy;
3159	uint32_t u32FeaturesEDX;
3160	uint32_t u32VendorEBX, u32VendorECX, u32VendorEDX, u32AMDFeatureEDX, u32AMDFeatureECX;
3161	uint64_t val;
3162
3163	ASMCpuId(0, &u32Dummy, &u32VendorEBX, &u32VendorECX, &u32VendorEDX);
3164	ASMCpuId(1, &u32Dummy, &u32Dummy, &u32FeaturesECX, &u32FeaturesEDX);
3165	/* Query AMD features. */
3166	ASMCpuId(0x80000001, &u32Dummy, &u32Dummy, &u32AMDFeatureECX, &u32AMDFeatureEDX);
3167
3168	if ( u32VendorEBX == X86_CPUID_VENDOR_INTEL_EBX
3169	&& u32VendorECX == X86_CPUID_VENDOR_INTEL_ECX
3170	&& u32VendorEDX == X86_CPUID_VENDOR_INTEL_EDX
3171	)
3172	{
3173	if ( (u32FeaturesECX & X86_CPUID_FEATURE_ECX_VMX)
3174	&& (u32FeaturesEDX & X86_CPUID_FEATURE_EDX_MSR)
3175	&& (u32FeaturesEDX & X86_CPUID_FEATURE_EDX_FXSR)
3176	)
3177	{
3178	val = ASMRdMsr(MSR_IA32_FEATURE_CONTROL);
3179	/*
3180	* Both the LOCK and VMXON bit must be set; otherwise VMXON will generate a #GP.
3181	* Once the lock bit is set, this MSR can no longer be modified.
3182	*/
3183	if ( (val & (MSR_IA32_FEATURE_CONTROL_VMXON\|MSR_IA32_FEATURE_CONTROL_LOCK))
3184	== (MSR_IA32_FEATURE_CONTROL_VMXON\|MSR_IA32_FEATURE_CONTROL_LOCK) /* enabled and locked */
3185	\|\| !(val & MSR_IA32_FEATURE_CONTROL_LOCK) /* not enabled, but not locked either */
3186	)
3187	{
3188	VMX_CAPABILITY vtCaps;
3189
3190	*pfCaps \|= SUPVTCAPS_VT_X;
3191
3192	vtCaps.u = ASMRdMsr(MSR_IA32_VMX_PROCBASED_CTLS);
3193	if (vtCaps.n.allowed1 & VMX_VMCS_CTRL_PROC_EXEC_USE_SECONDARY_EXEC_CTRL)
3194	{
3195	vtCaps.u = ASMRdMsr(MSR_IA32_VMX_PROCBASED_CTLS2);
3196	if (vtCaps.n.allowed1 & VMX_VMCS_CTRL_PROC_EXEC2_EPT)
3197	*pfCaps \|= SUPVTCAPS_NESTED_PAGING;
3198	}
3199	return VINF_SUCCESS;
3200	}
3201	return VERR_VMX_MSR_LOCKED_OR_DISABLED;
3202	}
3203	return VERR_VMX_NO_VMX;
3204	}
3205
3206	if ( u32VendorEBX == X86_CPUID_VENDOR_AMD_EBX
3207	&& u32VendorECX == X86_CPUID_VENDOR_AMD_ECX
3208	&& u32VendorEDX == X86_CPUID_VENDOR_AMD_EDX
3209	)
3210	{
3211	if ( (u32AMDFeatureECX & X86_CPUID_AMD_FEATURE_ECX_SVM)
3212	&& (u32FeaturesEDX & X86_CPUID_FEATURE_EDX_MSR)
3213	&& (u32FeaturesEDX & X86_CPUID_FEATURE_EDX_FXSR)
3214	)
3215	{
3216	/* Check if SVM is disabled */
3217	val = ASMRdMsr(MSR_K8_VM_CR);
3218	if (!(val & MSR_K8_VM_CR_SVM_DISABLE))
3219	{
3220	*pfCaps \|= SUPVTCAPS_AMD_V;
3221
3222	/* Query AMD features. */
3223	ASMCpuId(0x8000000A, &u32Dummy, &u32Dummy, &u32Dummy, &u32FeaturesEDX);
3224
3225	if (u32FeaturesEDX & AMD_CPUID_SVM_FEATURE_EDX_NESTED_PAGING)
3226	*pfCaps \|= SUPVTCAPS_NESTED_PAGING;
3227
3228	return VINF_SUCCESS;
3229	}
3230	return VERR_SVM_DISABLED;
3231	}
3232	return VERR_SVM_NO_SVM;
3233	}
3234	}
3235
3236	return VERR_UNSUPPORTED_CPU;
3237	}
3238
3239
3240	/**
3241	* (Re-)initializes the per-cpu structure prior to starting or resuming the GIP
3242	* updating.
3243	*
3244	* @param pGipCpu The per CPU structure for this CPU.
3245	* @param u64NanoTS The current time.
3246	*/
3247	static void supdrvGipReInitCpu(PSUPGIPCPU pGipCpu, uint64_t u64NanoTS)
3248	{
3249	pGipCpu->u64TSC = ASMReadTSC() - pGipCpu->u32UpdateIntervalTSC;
3250	pGipCpu->u64NanoTS = u64NanoTS;
3251	}
3252
3253
3254	/**
3255	* Set the current TSC and NanoTS value for the CPU.
3256	*
3257	* @param idCpu The CPU ID. Unused - we have to use the APIC ID.
3258	* @param pvUser1 Pointer to the ring-0 GIP mapping.
3259	* @param pvUser2 Pointer to the variable holding the current time.
3260	*/
3261	static DECLCALLBACK(void) supdrvGipReInitCpuCallback(RTCPUID idCpu, void pvUser1, void pvUser2)
3262	{
3263	PSUPGLOBALINFOPAGE pGip = (PSUPGLOBALINFOPAGE)pvUser1;
3264	unsigned iCpu = pGip->aiCpuFromApicId[ASMGetApicId()];
3265
3266	if (RT_LIKELY(iCpu < pGip->cCpus && pGip->aCPUs[iCpu].idCpu == idCpu))
3267	supdrvGipReInitCpu(&pGip->aCPUs[iCpu], (uint64_t )pvUser2);
3268
3269	NOREF(pvUser2);
3270	NOREF(idCpu);
3271	}
3272
3273
3274	/**
3275	* Maps the GIP into userspace and/or get the physical address of the GIP.
3276	*
3277	* @returns IPRT status code.
3278	* @param pSession Session to which the GIP mapping should belong.
3279	* @param ppGipR3 Where to store the address of the ring-3 mapping. (optional)
3280	* @param pHCPhysGip Where to store the physical address. (optional)
3281	*
3282	* @remark There is no reference counting on the mapping, so one call to this function
3283	* count globally as one reference. One call to SUPR0GipUnmap() is will unmap GIP
3284	* and remove the session as a GIP user.
3285	*/
3286	SUPR0DECL(int) SUPR0GipMap(PSUPDRVSESSION pSession, PRTR3PTR ppGipR3, PRTHCPHYS pHCPhysGip)
3287	{
3288	int rc;
3289	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
3290	RTR3PTR pGipR3 = NIL_RTR3PTR;
3291	RTHCPHYS HCPhys = NIL_RTHCPHYS;
3292	LogFlow(("SUPR0GipMap: pSession=%p ppGipR3=%p pHCPhysGip=%p\n", pSession, ppGipR3, pHCPhysGip));
3293
3294	/*
3295	* Validate
3296	*/
3297	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3298	AssertPtrNullReturn(ppGipR3, VERR_INVALID_POINTER);
3299	AssertPtrNullReturn(pHCPhysGip, VERR_INVALID_POINTER);
3300
3301	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3302	RTSemMutexRequest(pDevExt->mtxGip, RT_INDEFINITE_WAIT);
3303	#else
3304	RTSemFastMutexRequest(pDevExt->mtxGip);
3305	#endif
3306	if (pDevExt->pGip)
3307	{
3308	/*
3309	* Map it?
3310	*/
3311	rc = VINF_SUCCESS;
3312	if (ppGipR3)
3313	{
3314	if (pSession->GipMapObjR3 == NIL_RTR0MEMOBJ)
3315	rc = RTR0MemObjMapUser(&pSession->GipMapObjR3, pDevExt->GipMemObj, (RTR3PTR)-1, 0,
3316	RTMEM_PROT_READ, RTR0ProcHandleSelf());
3317	if (RT_SUCCESS(rc))
3318	pGipR3 = RTR0MemObjAddressR3(pSession->GipMapObjR3);
3319	}
3320
3321	/*
3322	* Get physical address.
3323	*/
3324	if (pHCPhysGip && RT_SUCCESS(rc))
3325	HCPhys = pDevExt->HCPhysGip;
3326
3327	/*
3328	* Reference globally.
3329	*/
3330	if (!pSession->fGipReferenced && RT_SUCCESS(rc))
3331	{
3332	pSession->fGipReferenced = 1;
3333	pDevExt->cGipUsers++;
3334	if (pDevExt->cGipUsers == 1)
3335	{
3336	PSUPGLOBALINFOPAGE pGipR0 = pDevExt->pGip;
3337	uint64_t u64NanoTS;
3338	uint32_t u32SystemResolution;
3339	unsigned i;
3340
3341	LogFlow(("SUPR0GipMap: Resumes GIP updating\n"));
3342
3343	/*
3344	* Try bump up the system timer resolution.
3345	* The more interrupts the better...
3346	*/
3347	if ( RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 976563 /* 1024 HZ */, &u32SystemResolution))
3348	\|\| RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 1000000 /* 1000 HZ */, &u32SystemResolution))
3349	\|\| RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 1953125 /* 512 HZ */, &u32SystemResolution))
3350	\|\| RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 2000000 /* 500 HZ */, &u32SystemResolution))
3351	)
3352	{
3353	Assert(RTTimerGetSystemGranularity() <= u32SystemResolution);
3354	pDevExt->u32SystemTimerGranularityGrant = u32SystemResolution;
3355	}
3356
3357	if (pGipR0->aCPUs[0].u32TransactionId != 2 /* not the first time */)
3358	{
3359	for (i = 0; i < RT_ELEMENTS(pGipR0->aCPUs); i++)
3360	ASMAtomicUoWriteU32(&pGipR0->aCPUs[i].u32TransactionId,
3361	(pGipR0->aCPUs[i].u32TransactionId + GIP_UPDATEHZ_RECALC_FREQ * 2)
3362	& ~(GIP_UPDATEHZ_RECALC_FREQ * 2 - 1));
3363	ASMAtomicWriteU64(&pGipR0->u64NanoTSLastUpdateHz, 0);
3364	}
3365
3366	u64NanoTS = RTTimeSystemNanoTS() - pGipR0->u32UpdateIntervalNS;
3367	if ( pGipR0->u32Mode == SUPGIPMODE_SYNC_TSC
3368	\|\| RTMpGetOnlineCount() == 1)
3369	supdrvGipReInitCpu(&pGipR0->aCPUs[0], u64NanoTS);
3370	else
3371	RTMpOnAll(supdrvGipReInitCpuCallback, pGipR0, &u64NanoTS);
3372
3373	#ifndef DO_NOT_START_GIP
3374	rc = RTTimerStart(pDevExt->pGipTimer, 0); AssertRC(rc);
3375	#endif
3376	rc = VINF_SUCCESS;
3377	}
3378	}
3379	}
3380	else
3381	{
3382	rc = VERR_GENERAL_FAILURE;
3383	Log(("SUPR0GipMap: GIP is not available!\n"));
3384	}
3385	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3386	RTSemMutexRelease(pDevExt->mtxGip);
3387	#else
3388	RTSemFastMutexRelease(pDevExt->mtxGip);
3389	#endif
3390
3391	/*
3392	* Write returns.
3393	*/
3394	if (pHCPhysGip)
3395	*pHCPhysGip = HCPhys;
3396	if (ppGipR3)
3397	*ppGipR3 = pGipR3;
3398
3399	#ifdef DEBUG_DARWIN_GIP
3400	OSDBGPRINT(("SUPR0GipMap: returns %d pHCPhysGip=%lx pGipR3=%p\n", rc, (unsigned long)HCPhys, (void )pGipR3));
3401	#else
3402	LogFlow(( "SUPR0GipMap: returns %d pHCPhysGip=%lx pGipR3=%p\n", rc, (unsigned long)HCPhys, (void )pGipR3));
3403	#endif
3404	return rc;
3405	}
3406
3407
3408	/**
3409	* Unmaps any user mapping of the GIP and terminates all GIP access
3410	* from this session.
3411	*
3412	* @returns IPRT status code.
3413	* @param pSession Session to which the GIP mapping should belong.
3414	*/
3415	SUPR0DECL(int) SUPR0GipUnmap(PSUPDRVSESSION pSession)
3416	{
3417	int rc = VINF_SUCCESS;
3418	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
3419	#ifdef DEBUG_DARWIN_GIP
3420	OSDBGPRINT(("SUPR0GipUnmap: pSession=%p pGip=%p GipMapObjR3=%p\n",
3421	pSession,
3422	pSession->GipMapObjR3 != NIL_RTR0MEMOBJ ? RTR0MemObjAddress(pSession->GipMapObjR3) : NULL,
3423	pSession->GipMapObjR3));
3424	#else
3425	LogFlow(("SUPR0GipUnmap: pSession=%p\n", pSession));
3426	#endif
3427	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3428
3429	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3430	RTSemMutexRequest(pDevExt->mtxGip, RT_INDEFINITE_WAIT);
3431	#else
3432	RTSemFastMutexRequest(pDevExt->mtxGip);
3433	#endif
3434
3435	/*
3436	* Unmap anything?
3437	*/
3438	if (pSession->GipMapObjR3 != NIL_RTR0MEMOBJ)
3439	{
3440	rc = RTR0MemObjFree(pSession->GipMapObjR3, false);
3441	AssertRC(rc);
3442	if (RT_SUCCESS(rc))
3443	pSession->GipMapObjR3 = NIL_RTR0MEMOBJ;
3444	}
3445
3446	/*
3447	* Dereference global GIP.
3448	*/
3449	if (pSession->fGipReferenced && !rc)
3450	{
3451	pSession->fGipReferenced = 0;
3452	if ( pDevExt->cGipUsers > 0
3453	&& !--pDevExt->cGipUsers)
3454	{
3455	LogFlow(("SUPR0GipUnmap: Suspends GIP updating\n"));
3456	#ifndef DO_NOT_START_GIP
3457	rc = RTTimerStop(pDevExt->pGipTimer); AssertRC(rc); rc = VINF_SUCCESS;
3458	#endif
3459
3460	if (pDevExt->u32SystemTimerGranularityGrant)
3461	{
3462	int rc2 = RTTimerReleaseSystemGranularity(pDevExt->u32SystemTimerGranularityGrant);
3463	AssertRC(rc2);
3464	pDevExt->u32SystemTimerGranularityGrant = 0;
3465	}
3466	}
3467	}
3468
3469	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3470	RTSemMutexRelease(pDevExt->mtxGip);
3471	#else
3472	RTSemFastMutexRelease(pDevExt->mtxGip);
3473	#endif
3474
3475	return rc;
3476	}
3477
3478
3479	/**
3480	* Gets the GIP pointer.
3481	*
3482	* @returns Pointer to the GIP or NULL.
3483	*/
3484	SUPDECL(PSUPGLOBALINFOPAGE) SUPGetGIP(void)
3485	{
3486	return g_pSUPGlobalInfoPage;
3487	}
3488
3489
3490	/**
3491	* Register a component factory with the support driver.
3492	*
3493	* This is currently restricted to kernel sessions only.
3494	*
3495	* @returns VBox status code.
3496	* @retval VINF_SUCCESS on success.
3497	* @retval VERR_NO_MEMORY if we're out of memory.
3498	* @retval VERR_ALREADY_EXISTS if the factory has already been registered.
3499	* @retval VERR_ACCESS_DENIED if it isn't a kernel session.
3500	* @retval VERR_INVALID_PARAMETER on invalid parameter.
3501	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
3502	*
3503	* @param pSession The SUPDRV session (must be a ring-0 session).
3504	* @param pFactory Pointer to the component factory registration structure.
3505	*
3506	* @remarks This interface is also available via SUPR0IdcComponentRegisterFactory.
3507	*/
3508	SUPR0DECL(int) SUPR0ComponentRegisterFactory(PSUPDRVSESSION pSession, PCSUPDRVFACTORY pFactory)
3509	{
3510	PSUPDRVFACTORYREG pNewReg;
3511	const char *psz;
3512	int rc;
3513
3514	/*
3515	* Validate parameters.
3516	*/
3517	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3518	AssertReturn(pSession->R0Process == NIL_RTR0PROCESS, VERR_ACCESS_DENIED);
3519	AssertPtrReturn(pFactory, VERR_INVALID_POINTER);
3520	AssertPtrReturn(pFactory->pfnQueryFactoryInterface, VERR_INVALID_POINTER);
3521	psz = RTStrEnd(pFactory->szName, sizeof(pFactory->szName));
3522	AssertReturn(psz, VERR_INVALID_PARAMETER);
3523
3524	/*
3525	* Allocate and initialize a new registration structure.
3526	*/
3527	pNewReg = (PSUPDRVFACTORYREG)RTMemAlloc(sizeof(SUPDRVFACTORYREG));
3528	if (pNewReg)
3529	{
3530	pNewReg->pNext = NULL;
3531	pNewReg->pFactory = pFactory;
3532	pNewReg->pSession = pSession;
3533	pNewReg->cchName = psz - &pFactory->szName[0];
3534
3535	/*
3536	* Add it to the tail of the list after checking for prior registration.
3537	*/
3538	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
3539	if (RT_SUCCESS(rc))
3540	{
3541	PSUPDRVFACTORYREG pPrev = NULL;
3542	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
3543	while (pCur && pCur->pFactory != pFactory)
3544	{
3545	pPrev = pCur;
3546	pCur = pCur->pNext;
3547	}
3548	if (!pCur)
3549	{
3550	if (pPrev)
3551	pPrev->pNext = pNewReg;
3552	else
3553	pSession->pDevExt->pComponentFactoryHead = pNewReg;
3554	rc = VINF_SUCCESS;
3555	}
3556	else
3557	rc = VERR_ALREADY_EXISTS;
3558
3559	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
3560	}
3561
3562	if (RT_FAILURE(rc))
3563	RTMemFree(pNewReg);
3564	}
3565	else
3566	rc = VERR_NO_MEMORY;
3567	return rc;
3568	}
3569
3570
3571	/**
3572	* Deregister a component factory.
3573	*
3574	* @returns VBox status code.
3575	* @retval VINF_SUCCESS on success.
3576	* @retval VERR_NOT_FOUND if the factory wasn't registered.
3577	* @retval VERR_ACCESS_DENIED if it isn't a kernel session.
3578	* @retval VERR_INVALID_PARAMETER on invalid parameter.
3579	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
3580	*
3581	* @param pSession The SUPDRV session (must be a ring-0 session).
3582	* @param pFactory Pointer to the component factory registration structure
3583	* previously passed SUPR0ComponentRegisterFactory().
3584	*
3585	* @remarks This interface is also available via SUPR0IdcComponentDeregisterFactory.
3586	*/
3587	SUPR0DECL(int) SUPR0ComponentDeregisterFactory(PSUPDRVSESSION pSession, PCSUPDRVFACTORY pFactory)
3588	{
3589	int rc;
3590
3591	/*
3592	* Validate parameters.
3593	*/
3594	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3595	AssertReturn(pSession->R0Process == NIL_RTR0PROCESS, VERR_ACCESS_DENIED);
3596	AssertPtrReturn(pFactory, VERR_INVALID_POINTER);
3597
3598	/*
3599	* Take the lock and look for the registration record.
3600	*/
3601	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
3602	if (RT_SUCCESS(rc))
3603	{
3604	PSUPDRVFACTORYREG pPrev = NULL;
3605	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
3606	while (pCur && pCur->pFactory != pFactory)
3607	{
3608	pPrev = pCur;
3609	pCur = pCur->pNext;
3610	}
3611	if (pCur)
3612	{
3613	if (!pPrev)
3614	pSession->pDevExt->pComponentFactoryHead = pCur->pNext;
3615	else
3616	pPrev->pNext = pCur->pNext;
3617
3618	pCur->pNext = NULL;
3619	pCur->pFactory = NULL;
3620	pCur->pSession = NULL;
3621	rc = VINF_SUCCESS;
3622	}
3623	else
3624	rc = VERR_NOT_FOUND;
3625
3626	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
3627
3628	RTMemFree(pCur);
3629	}
3630	return rc;
3631	}
3632
3633
3634	/**
3635	* Queries a component factory.
3636	*
3637	* @returns VBox status code.
3638	* @retval VERR_INVALID_PARAMETER on invalid parameter.
3639	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
3640	* @retval VERR_SUPDRV_COMPONENT_NOT_FOUND if the component factory wasn't found.
3641	* @retval VERR_SUPDRV_INTERFACE_NOT_SUPPORTED if the interface wasn't supported.
3642	*
3643	* @param pSession The SUPDRV session.
3644	* @param pszName The name of the component factory.
3645	* @param pszInterfaceUuid The UUID of the factory interface (stringified).
3646	* @param ppvFactoryIf Where to store the factory interface.
3647	*/
3648	SUPR0DECL(int) SUPR0ComponentQueryFactory(PSUPDRVSESSION pSession, const char pszName, const char pszInterfaceUuid, void **ppvFactoryIf)
3649	{
3650	const char *pszEnd;
3651	size_t cchName;
3652	int rc;
3653
3654	/*
3655	* Validate parameters.
3656	*/
3657	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3658
3659	AssertPtrReturn(pszName, VERR_INVALID_POINTER);
3660	pszEnd = RTStrEnd(pszName, RT_SIZEOFMEMB(SUPDRVFACTORY, szName));
3661	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
3662	cchName = pszEnd - pszName;
3663
3664	AssertPtrReturn(pszInterfaceUuid, VERR_INVALID_POINTER);
3665	pszEnd = RTStrEnd(pszInterfaceUuid, RTUUID_STR_LENGTH);
3666	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
3667
3668	AssertPtrReturn(ppvFactoryIf, VERR_INVALID_POINTER);
3669	*ppvFactoryIf = NULL;
3670
3671	/*
3672	* Take the lock and try all factories by this name.
3673	*/
3674	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
3675	if (RT_SUCCESS(rc))
3676	{
3677	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
3678	rc = VERR_SUPDRV_COMPONENT_NOT_FOUND;
3679	while (pCur)
3680	{
3681	if ( pCur->cchName == cchName
3682	&& !memcmp(pCur->pFactory->szName, pszName, cchName))
3683	{
3684	void *pvFactory = pCur->pFactory->pfnQueryFactoryInterface(pCur->pFactory, pSession, pszInterfaceUuid);
3685	if (pvFactory)
3686	{
3687	*ppvFactoryIf = pvFactory;
3688	rc = VINF_SUCCESS;
3689	break;
3690	}
3691	rc = VERR_SUPDRV_INTERFACE_NOT_SUPPORTED;
3692	}
3693
3694	/* next */
3695	pCur = pCur->pNext;
3696	}
3697
3698	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
3699	}
3700	return rc;
3701	}
3702
3703
3704	/**
3705	* Adds a memory object to the session.
3706	*
3707	* @returns IPRT status code.
3708	* @param pMem Memory tracking structure containing the
3709	* information to track.
3710	* @param pSession The session.
3711	*/
3712	static int supdrvMemAdd(PSUPDRVMEMREF pMem, PSUPDRVSESSION pSession)
3713	{
3714	PSUPDRVBUNDLE pBundle;
3715
3716	/*
3717	* Find free entry and record the allocation.
3718	*/
3719	RTSpinlockAcquire(pSession->Spinlock);
3720	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
3721	{
3722	if (pBundle->cUsed < RT_ELEMENTS(pBundle->aMem))
3723	{
3724	unsigned i;
3725	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
3726	{
3727	if (pBundle->aMem[i].MemObj == NIL_RTR0MEMOBJ)
3728	{
3729	pBundle->cUsed++;
3730	pBundle->aMem[i] = *pMem;
3731	RTSpinlockRelease(pSession->Spinlock);
3732	return VINF_SUCCESS;
3733	}
3734	}
3735	AssertFailed(); /* !!this can't be happening!!! */
3736	}
3737	}
3738	RTSpinlockRelease(pSession->Spinlock);
3739
3740	/*
3741	* Need to allocate a new bundle.
3742	* Insert into the last entry in the bundle.
3743	*/
3744	pBundle = (PSUPDRVBUNDLE)RTMemAllocZ(sizeof(*pBundle));
3745	if (!pBundle)
3746	return VERR_NO_MEMORY;
3747
3748	/* take last entry. */
3749	pBundle->cUsed++;
3750	pBundle->aMem[RT_ELEMENTS(pBundle->aMem) - 1] = *pMem;
3751
3752	/* insert into list. */
3753	RTSpinlockAcquire(pSession->Spinlock);
3754	pBundle->pNext = pSession->Bundle.pNext;
3755	pSession->Bundle.pNext = pBundle;
3756	RTSpinlockRelease(pSession->Spinlock);
3757
3758	return VINF_SUCCESS;
3759	}
3760
3761
3762	/**
3763	* Releases a memory object referenced by pointer and type.
3764	*
3765	* @returns IPRT status code.
3766	* @param pSession Session data.
3767	* @param uPtr Pointer to memory. This is matched against both the R0 and R3 addresses.
3768	* @param eType Memory type.
3769	*/
3770	static int supdrvMemRelease(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, SUPDRVMEMREFTYPE eType)
3771	{
3772	PSUPDRVBUNDLE pBundle;
3773
3774	/*
3775	* Validate input.
3776	*/
3777	if (!uPtr)
3778	{
3779	Log(("Illegal address %p\n", (void *)uPtr));
3780	return VERR_INVALID_PARAMETER;
3781	}
3782
3783	/*
3784	* Search for the address.
3785	*/
3786	RTSpinlockAcquire(pSession->Spinlock);
3787	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
3788	{
3789	if (pBundle->cUsed > 0)
3790	{
3791	unsigned i;
3792	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
3793	{
3794	if ( pBundle->aMem[i].eType == eType
3795	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
3796	&& ( (RTHCUINTPTR)RTR0MemObjAddress(pBundle->aMem[i].MemObj) == uPtr
3797	\|\| ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
3798	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == uPtr))
3799	)
3800	{
3801	/* Make a copy of it and release it outside the spinlock. */
3802	SUPDRVMEMREF Mem = pBundle->aMem[i];
3803	pBundle->aMem[i].eType = MEMREF_TYPE_UNUSED;
3804	pBundle->aMem[i].MemObj = NIL_RTR0MEMOBJ;
3805	pBundle->aMem[i].MapObjR3 = NIL_RTR0MEMOBJ;
3806	RTSpinlockRelease(pSession->Spinlock);
3807
3808	if (Mem.MapObjR3 != NIL_RTR0MEMOBJ)
3809	{
3810	int rc = RTR0MemObjFree(Mem.MapObjR3, false);
3811	AssertRC(rc); /** @todo figure out how to handle this. */
3812	}
3813	if (Mem.MemObj != NIL_RTR0MEMOBJ)
3814	{
3815	int rc = RTR0MemObjFree(Mem.MemObj, true /* fFreeMappings */);
3816	AssertRC(rc); /** @todo figure out how to handle this. */
3817	}
3818	return VINF_SUCCESS;
3819	}
3820	}
3821	}
3822	}
3823	RTSpinlockRelease(pSession->Spinlock);
3824	Log(("Failed to find %p!!! (eType=%d)\n", (void *)uPtr, eType));
3825	return VERR_INVALID_PARAMETER;
3826	}
3827
3828
3829	/**
3830	* Opens an image. If it's the first time it's opened the call must upload
3831	* the bits using the supdrvIOCtl_LdrLoad() / SUPDRV_IOCTL_LDR_LOAD function.
3832	*
3833	* This is the 1st step of the loading.
3834	*
3835	* @returns IPRT status code.
3836	* @param pDevExt Device globals.
3837	* @param pSession Session data.
3838	* @param pReq The open request.
3839	*/
3840	static int supdrvIOCtl_LdrOpen(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDROPEN pReq)
3841	{
3842	int rc;
3843	PSUPDRVLDRIMAGE pImage;
3844	void *pv;
3845	size_t cchName = strlen(pReq->u.In.szName); /* (caller checked < 32). */
3846	LogFlow(("supdrvIOCtl_LdrOpen: szName=%s cbImageWithTabs=%d\n", pReq->u.In.szName, pReq->u.In.cbImageWithTabs));
3847
3848	/*
3849	* Check if we got an instance of the image already.
3850	*/
3851	supdrvLdrLock(pDevExt);
3852	for (pImage = pDevExt->pLdrImages; pImage; pImage = pImage->pNext)
3853	{
3854	if ( pImage->szName[cchName] == '\0'
3855	&& !memcmp(pImage->szName, pReq->u.In.szName, cchName))
3856	{
3857	/** @todo check cbImageBits and cbImageWithTabs here, if they differs that indicates that the images are different. */
3858	pImage->cUsage++;
3859	pReq->u.Out.pvImageBase = pImage->pvImage;
3860	pReq->u.Out.fNeedsLoading = pImage->uState == SUP_IOCTL_LDR_OPEN;
3861	pReq->u.Out.fNativeLoader = pImage->fNative;
3862	supdrvLdrAddUsage(pSession, pImage);
3863	supdrvLdrUnlock(pDevExt);
3864	return VINF_SUCCESS;
3865	}
3866	}
3867	/* (not found - add it!) */
3868
3869	/*
3870	* Allocate memory.
3871	*/
3872	pv = RTMemAlloc(RT_OFFSETOF(SUPDRVLDRIMAGE, szName[cchName + 1]));
3873	if (!pv)
3874	{
3875	supdrvLdrUnlock(pDevExt);
3876	Log(("supdrvIOCtl_LdrOpen: RTMemAlloc() failed\n"));
3877	return /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_2;
3878	}
3879
3880	/*
3881	* Setup and link in the LDR stuff.
3882	*/
3883	pImage = (PSUPDRVLDRIMAGE)pv;
3884	pImage->pvImage = NULL;
3885	pImage->pvImageAlloc = NULL;
3886	pImage->cbImageWithTabs = pReq->u.In.cbImageWithTabs;
3887	pImage->cbImageBits = pReq->u.In.cbImageBits;
3888	pImage->cSymbols = 0;
3889	pImage->paSymbols = NULL;
3890	pImage->pachStrTab = NULL;
3891	pImage->cbStrTab = 0;
3892	pImage->pfnModuleInit = NULL;
3893	pImage->pfnModuleTerm = NULL;
3894	pImage->pfnServiceReqHandler = NULL;
3895	pImage->uState = SUP_IOCTL_LDR_OPEN;
3896	pImage->cUsage = 1;
3897	pImage->pDevExt = pDevExt;
3898	memcpy(pImage->szName, pReq->u.In.szName, cchName + 1);
3899
3900	/*
3901	* Try load it using the native loader, if that isn't supported, fall back
3902	* on the older method.
3903	*/
3904	pImage->fNative = true;
3905	rc = supdrvOSLdrOpen(pDevExt, pImage, pReq->u.In.szFilename);
3906	if (rc == VERR_NOT_SUPPORTED)
3907	{
3908	pImage->pvImageAlloc = RTMemExecAlloc(pImage->cbImageBits + 31);
3909	pImage->pvImage = RT_ALIGN_P(pImage->pvImageAlloc, 32);
3910	pImage->fNative = false;
3911	rc = pImage->pvImageAlloc ? VINF_SUCCESS : VERR_NO_EXEC_MEMORY;
3912	}
3913	if (RT_FAILURE(rc))
3914	{
3915	supdrvLdrUnlock(pDevExt);
3916	RTMemFree(pImage);
3917	Log(("supdrvIOCtl_LdrOpen(%s): failed - %Rrc\n", pReq->u.In.szName, rc));
3918	return rc;
3919	}
3920	Assert(VALID_PTR(pImage->pvImage) \|\| RT_FAILURE(rc));
3921
3922	/*
3923	* Link it.
3924	*/
3925	pImage->pNext = pDevExt->pLdrImages;
3926	pDevExt->pLdrImages = pImage;
3927
3928	supdrvLdrAddUsage(pSession, pImage);
3929
3930	pReq->u.Out.pvImageBase = pImage->pvImage;
3931	pReq->u.Out.fNeedsLoading = true;
3932	pReq->u.Out.fNativeLoader = pImage->fNative;
3933	supdrvLdrUnlock(pDevExt);
3934
3935	#if defined(RT_OS_WINDOWS) && defined(DEBUG)
3936	SUPR0Printf("VBoxDrv: windbg> .reload /f %s=%#p\n", pImage->szName, pImage->pvImage);
3937	#endif
3938	return VINF_SUCCESS;
3939	}
3940
3941
3942	/**
3943	* Worker that validates a pointer to an image entrypoint.
3944	*
3945	* @returns IPRT status code.
3946	* @param pDevExt The device globals.
3947	* @param pImage The loader image.
3948	* @param pv The pointer into the image.
3949	* @param fMayBeNull Whether it may be NULL.
3950	* @param pszWhat What is this entrypoint? (for logging)
3951	* @param pbImageBits The image bits prepared by ring-3.
3952	*
3953	* @remarks Will leave the lock on failure.
3954	*/
3955	static int supdrvLdrValidatePointer(PSUPDRVDEVEXT pDevExt, PSUPDRVLDRIMAGE pImage, void *pv,
3956	bool fMayBeNull, const uint8_t pbImageBits, const char pszWhat)
3957	{
3958	if (!fMayBeNull \|\| pv)
3959	{
3960	if ((uintptr_t)pv - (uintptr_t)pImage->pvImage >= pImage->cbImageBits)
3961	{
3962	supdrvLdrUnlock(pDevExt);
3963	Log(("Out of range (%p LB %#x): %s=%p\n", pImage->pvImage, pImage->cbImageBits, pszWhat, pv));
3964	return VERR_INVALID_PARAMETER;
3965	}
3966
3967	if (pImage->fNative)
3968	{
3969	int rc = supdrvOSLdrValidatePointer(pDevExt, pImage, pv, pbImageBits);
3970	if (RT_FAILURE(rc))
3971	{
3972	supdrvLdrUnlock(pDevExt);
3973	Log(("Bad entry point address: %s=%p (rc=%Rrc)\n", pszWhat, pv, rc));
3974	return rc;
3975	}
3976	}
3977	}
3978	return VINF_SUCCESS;
3979	}
3980
3981
3982	/**
3983	* Loads the image bits.
3984	*
3985	* This is the 2nd step of the loading.
3986	*
3987	* @returns IPRT status code.
3988	* @param pDevExt Device globals.
3989	* @param pSession Session data.
3990	* @param pReq The request.
3991	*/
3992	static int supdrvIOCtl_LdrLoad(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRLOAD pReq)
3993	{
3994	PSUPDRVLDRUSAGE pUsage;
3995	PSUPDRVLDRIMAGE pImage;
3996	int rc;
3997	LogFlow(("supdrvIOCtl_LdrLoad: pvImageBase=%p cbImageWithBits=%d\n", pReq->u.In.pvImageBase, pReq->u.In.cbImageWithTabs));
3998
3999	/*
4000	* Find the ldr image.
4001	*/
4002	supdrvLdrLock(pDevExt);
4003	pUsage = pSession->pLdrUsage;
4004	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
4005	pUsage = pUsage->pNext;
4006	if (!pUsage)
4007	{
4008	supdrvLdrUnlock(pDevExt);
4009	Log(("SUP_IOCTL_LDR_LOAD: couldn't find image!\n"));
4010	return VERR_INVALID_HANDLE;
4011	}
4012	pImage = pUsage->pImage;
4013
4014	/*
4015	* Validate input.
4016	*/
4017	if ( pImage->cbImageWithTabs != pReq->u.In.cbImageWithTabs
4018	\|\| pImage->cbImageBits != pReq->u.In.cbImageBits)
4019	{
4020	supdrvLdrUnlock(pDevExt);
4021	Log(("SUP_IOCTL_LDR_LOAD: image size mismatch!! %d(prep) != %d(load) or %d != %d\n",
4022	pImage->cbImageWithTabs, pReq->u.In.cbImageWithTabs, pImage->cbImageBits, pReq->u.In.cbImageBits));
4023	return VERR_INVALID_HANDLE;
4024	}
4025
4026	if (pImage->uState != SUP_IOCTL_LDR_OPEN)
4027	{
4028	unsigned uState = pImage->uState;
4029	supdrvLdrUnlock(pDevExt);
4030	if (uState != SUP_IOCTL_LDR_LOAD)
4031	AssertMsgFailed(("SUP_IOCTL_LDR_LOAD: invalid image state %d (%#x)!\n", uState, uState));
4032	return VERR_ALREADY_LOADED;
4033	}
4034
4035	switch (pReq->u.In.eEPType)
4036	{
4037	case SUPLDRLOADEP_NOTHING:
4038	break;
4039
4040	case SUPLDRLOADEP_VMMR0:
4041	rc = supdrvLdrValidatePointer( pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0, false, pReq->u.In.abImage, "pvVMMR0");
4042	if (RT_SUCCESS(rc))
4043	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0EntryInt, false, pReq->u.In.abImage, "pvVMMR0EntryInt");
4044	if (RT_SUCCESS(rc))
4045	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0EntryFast, false, pReq->u.In.abImage, "pvVMMR0EntryFast");
4046	if (RT_SUCCESS(rc))
4047	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0EntryEx, false, pReq->u.In.abImage, "pvVMMR0EntryEx");
4048	if (RT_FAILURE(rc))
4049	return rc;
4050	break;
4051
4052	case SUPLDRLOADEP_SERVICE:
4053	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.Service.pfnServiceReq, false, pReq->u.In.abImage, "pfnServiceReq");
4054	if (RT_FAILURE(rc))
4055	return rc;
4056	if ( pReq->u.In.EP.Service.apvReserved[0] != NIL_RTR0PTR
4057	\|\| pReq->u.In.EP.Service.apvReserved[1] != NIL_RTR0PTR
4058	\|\| pReq->u.In.EP.Service.apvReserved[2] != NIL_RTR0PTR)
4059	{
4060	supdrvLdrUnlock(pDevExt);
4061	Log(("Out of range (%p LB %#x): apvReserved={%p,%p,%p} MBZ!\n",
4062	pImage->pvImage, pReq->u.In.cbImageWithTabs,
4063	pReq->u.In.EP.Service.apvReserved[0],
4064	pReq->u.In.EP.Service.apvReserved[1],
4065	pReq->u.In.EP.Service.apvReserved[2]));
4066	return VERR_INVALID_PARAMETER;
4067	}
4068	break;
4069
4070	default:
4071	supdrvLdrUnlock(pDevExt);
4072	Log(("Invalid eEPType=%d\n", pReq->u.In.eEPType));
4073	return VERR_INVALID_PARAMETER;
4074	}
4075
4076	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.pfnModuleInit, true, pReq->u.In.abImage, "pfnModuleInit");
4077	if (RT_FAILURE(rc))
4078	return rc;
4079	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.pfnModuleTerm, true, pReq->u.In.abImage, "pfnModuleTerm");
4080	if (RT_FAILURE(rc))
4081	return rc;
4082
4083	/*
4084	* Allocate and copy the tables.
4085	* (No need to do try/except as this is a buffered request.)
4086	*/
4087	pImage->cbStrTab = pReq->u.In.cbStrTab;
4088	if (pImage->cbStrTab)
4089	{
4090	pImage->pachStrTab = (char *)RTMemAlloc(pImage->cbStrTab);
4091	if (pImage->pachStrTab)
4092	memcpy(pImage->pachStrTab, &pReq->u.In.abImage[pReq->u.In.offStrTab], pImage->cbStrTab);
4093	else
4094	rc = /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_3;
4095	}
4096
4097	pImage->cSymbols = pReq->u.In.cSymbols;
4098	if (RT_SUCCESS(rc) && pImage->cSymbols)
4099	{
4100	size_t cbSymbols = pImage->cSymbols * sizeof(SUPLDRSYM);
4101	pImage->paSymbols = (PSUPLDRSYM)RTMemAlloc(cbSymbols);
4102	if (pImage->paSymbols)
4103	memcpy(pImage->paSymbols, &pReq->u.In.abImage[pReq->u.In.offSymbols], cbSymbols);
4104	else
4105	rc = /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_4;
4106	}
4107
4108	/*
4109	* Copy the bits / complete native loading.
4110	*/
4111	if (RT_SUCCESS(rc))
4112	{
4113	pImage->uState = SUP_IOCTL_LDR_LOAD;
4114	pImage->pfnModuleInit = pReq->u.In.pfnModuleInit;
4115	pImage->pfnModuleTerm = pReq->u.In.pfnModuleTerm;
4116
4117	if (pImage->fNative)
4118	rc = supdrvOSLdrLoad(pDevExt, pImage, pReq->u.In.abImage, pReq);
4119	else
4120	{
4121	memcpy(pImage->pvImage, &pReq->u.In.abImage[0], pImage->cbImageBits);
4122	SUPR0Printf("vboxdrv: Loaded '%s' at %p\n", pImage->szName, pImage->pvImage);
4123	}
4124	}
4125
4126	/*
4127	* Update any entry points.
4128	*/
4129	if (RT_SUCCESS(rc))
4130	{
4131	switch (pReq->u.In.eEPType)
4132	{
4133	default:
4134	case SUPLDRLOADEP_NOTHING:
4135	rc = VINF_SUCCESS;
4136	break;
4137	case SUPLDRLOADEP_VMMR0:
4138	rc = supdrvLdrSetVMMR0EPs(pDevExt, pReq->u.In.EP.VMMR0.pvVMMR0, pReq->u.In.EP.VMMR0.pvVMMR0EntryInt,
4139	pReq->u.In.EP.VMMR0.pvVMMR0EntryFast, pReq->u.In.EP.VMMR0.pvVMMR0EntryEx);
4140	break;
4141	case SUPLDRLOADEP_SERVICE:
4142	pImage->pfnServiceReqHandler = pReq->u.In.EP.Service.pfnServiceReq;
4143	rc = VINF_SUCCESS;
4144	break;
4145	}
4146	}
4147
4148	/*
4149	* On success call the module initialization.
4150	*/
4151	LogFlow(("supdrvIOCtl_LdrLoad: pfnModuleInit=%p\n", pImage->pfnModuleInit));
4152	if (RT_SUCCESS(rc) && pImage->pfnModuleInit)
4153	{
4154	Log(("supdrvIOCtl_LdrLoad: calling pfnModuleInit=%p\n", pImage->pfnModuleInit));
4155	pDevExt->pLdrInitImage = pImage;
4156	pDevExt->hLdrInitThread = RTThreadNativeSelf();
4157	rc = pImage->pfnModuleInit(pImage);
4158	pDevExt->pLdrInitImage = NULL;
4159	pDevExt->hLdrInitThread = NIL_RTNATIVETHREAD;
4160	if (RT_FAILURE(rc) && pDevExt->pvVMMR0 == pImage->pvImage)
4161	supdrvLdrUnsetVMMR0EPs(pDevExt);
4162	}
4163
4164	if (RT_FAILURE(rc))
4165	{
4166	/* Inform the tracing component in case ModuleInit registered TPs. */
4167	supdrvTracerModuleUnloading(pDevExt, pImage);
4168
4169	pImage->uState = SUP_IOCTL_LDR_OPEN;
4170	pImage->pfnModuleInit = NULL;
4171	pImage->pfnModuleTerm = NULL;
4172	pImage->pfnServiceReqHandler= NULL;
4173	pImage->cbStrTab = 0;
4174	RTMemFree(pImage->pachStrTab);
4175	pImage->pachStrTab = NULL;
4176	RTMemFree(pImage->paSymbols);
4177	pImage->paSymbols = NULL;
4178	pImage->cSymbols = 0;
4179	}
4180
4181	supdrvLdrUnlock(pDevExt);
4182	return rc;
4183	}
4184
4185
4186	/**
4187	* Frees a previously loaded (prep'ed) image.
4188	*
4189	* @returns IPRT status code.
4190	* @param pDevExt Device globals.
4191	* @param pSession Session data.
4192	* @param pReq The request.
4193	*/
4194	static int supdrvIOCtl_LdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRFREE pReq)
4195	{
4196	int rc;
4197	PSUPDRVLDRUSAGE pUsagePrev;
4198	PSUPDRVLDRUSAGE pUsage;
4199	PSUPDRVLDRIMAGE pImage;
4200	LogFlow(("supdrvIOCtl_LdrFree: pvImageBase=%p\n", pReq->u.In.pvImageBase));
4201
4202	/*
4203	* Find the ldr image.
4204	*/
4205	supdrvLdrLock(pDevExt);
4206	pUsagePrev = NULL;
4207	pUsage = pSession->pLdrUsage;
4208	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
4209	{
4210	pUsagePrev = pUsage;
4211	pUsage = pUsage->pNext;
4212	}
4213	if (!pUsage)
4214	{
4215	supdrvLdrUnlock(pDevExt);
4216	Log(("SUP_IOCTL_LDR_FREE: couldn't find image!\n"));
4217	return VERR_INVALID_HANDLE;
4218	}
4219
4220	/*
4221	* Check if we can remove anything.
4222	*/
4223	rc = VINF_SUCCESS;
4224	pImage = pUsage->pImage;
4225	if (pImage->cUsage <= 1 \|\| pUsage->cUsage <= 1)
4226	{
4227	/*
4228	* Check if there are any objects with destructors in the image, if
4229	* so leave it for the session cleanup routine so we get a chance to
4230	* clean things up in the right order and not leave them all dangling.
4231	*/
4232	RTSpinlockAcquire(pDevExt->Spinlock);
4233	if (pImage->cUsage <= 1)
4234	{
4235	PSUPDRVOBJ pObj;
4236	for (pObj = pDevExt->pObjs; pObj; pObj = pObj->pNext)
4237	if (RT_UNLIKELY((uintptr_t)pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImageBits))
4238	{
4239	rc = VERR_DANGLING_OBJECTS;
4240	break;
4241	}
4242	}
4243	else
4244	{
4245	PSUPDRVUSAGE pGenUsage;
4246	for (pGenUsage = pSession->pUsage; pGenUsage; pGenUsage = pGenUsage->pNext)
4247	if (RT_UNLIKELY((uintptr_t)pGenUsage->pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImageBits))
4248	{
4249	rc = VERR_DANGLING_OBJECTS;
4250	break;
4251	}
4252	}
4253	RTSpinlockRelease(pDevExt->Spinlock);
4254	if (rc == VINF_SUCCESS)
4255	{
4256	/* unlink it */
4257	if (pUsagePrev)
4258	pUsagePrev->pNext = pUsage->pNext;
4259	else
4260	pSession->pLdrUsage = pUsage->pNext;
4261
4262	/* free it */
4263	pUsage->pImage = NULL;
4264	pUsage->pNext = NULL;
4265	RTMemFree(pUsage);
4266
4267	/*
4268	* Dereference the image.
4269	*/
4270	if (pImage->cUsage <= 1)
4271	supdrvLdrFree(pDevExt, pImage);
4272	else
4273	pImage->cUsage--;
4274	}
4275	else
4276	{
4277	Log(("supdrvIOCtl_LdrFree: Dangling objects in %p/%s!\n", pImage->pvImage, pImage->szName));
4278	rc = VINF_SUCCESS; /** @todo BRANCH-2.1: remove this after branching. */
4279	}
4280	}
4281	else
4282	{
4283	/*
4284	* Dereference both image and usage.
4285	*/
4286	pImage->cUsage--;
4287	pUsage->cUsage--;
4288	}
4289
4290	supdrvLdrUnlock(pDevExt);
4291	return rc;
4292	}
4293
4294
4295	/**
4296	* Gets the address of a symbol in an open image.
4297	*
4298	* @returns IPRT status code.
4299	* @param pDevExt Device globals.
4300	* @param pSession Session data.
4301	* @param pReq The request buffer.
4302	*/
4303	static int supdrvIOCtl_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRGETSYMBOL pReq)
4304	{
4305	PSUPDRVLDRIMAGE pImage;
4306	PSUPDRVLDRUSAGE pUsage;
4307	uint32_t i;
4308	PSUPLDRSYM paSyms;
4309	const char *pchStrings;
4310	const size_t cbSymbol = strlen(pReq->u.In.szSymbol) + 1;
4311	void *pvSymbol = NULL;
4312	int rc = VERR_GENERAL_FAILURE;
4313	Log3(("supdrvIOCtl_LdrGetSymbol: pvImageBase=%p szSymbol=\"%s\"\n", pReq->u.In.pvImageBase, pReq->u.In.szSymbol));
4314
4315	/*
4316	* Find the ldr image.
4317	*/
4318	supdrvLdrLock(pDevExt);
4319	pUsage = pSession->pLdrUsage;
4320	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
4321	pUsage = pUsage->pNext;
4322	if (!pUsage)
4323	{
4324	supdrvLdrUnlock(pDevExt);
4325	Log(("SUP_IOCTL_LDR_GET_SYMBOL: couldn't find image!\n"));
4326	return VERR_INVALID_HANDLE;
4327	}
4328	pImage = pUsage->pImage;
4329	if (pImage->uState != SUP_IOCTL_LDR_LOAD)
4330	{
4331	unsigned uState = pImage->uState;
4332	supdrvLdrUnlock(pDevExt);
4333	Log(("SUP_IOCTL_LDR_GET_SYMBOL: invalid image state %d (%#x)!\n", uState, uState)); NOREF(uState);
4334	return VERR_ALREADY_LOADED;
4335	}
4336
4337	/*
4338	* Search the symbol strings.
4339	*
4340	* Note! The int32_t is for native loading on solaris where the data
4341	* and text segments are in very different places.
4342	*/
4343	pchStrings = pImage->pachStrTab;
4344	paSyms = pImage->paSymbols;
4345	for (i = 0; i < pImage->cSymbols; i++)
4346	{
4347	if ( paSyms[i].offName + cbSymbol <= pImage->cbStrTab
4348	&& !memcmp(pchStrings + paSyms[i].offName, pReq->u.In.szSymbol, cbSymbol))
4349	{
4350	pvSymbol = (uint8_t *)pImage->pvImage + (int32_t)paSyms[i].offSymbol;
4351	rc = VINF_SUCCESS;
4352	break;
4353	}
4354	}
4355	supdrvLdrUnlock(pDevExt);
4356	pReq->u.Out.pvSymbol = pvSymbol;
4357	return rc;
4358	}
4359
4360
4361	/**
4362	* Gets the address of a symbol in an open image or the support driver.
4363	*
4364	* @returns VINF_SUCCESS on success.
4365	* @returns
4366	* @param pDevExt Device globals.
4367	* @param pSession Session data.
4368	* @param pReq The request buffer.
4369	*/
4370	static int supdrvIDC_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVIDCREQGETSYM pReq)
4371	{
4372	int rc = VINF_SUCCESS;
4373	const char *pszSymbol = pReq->u.In.pszSymbol;
4374	const char *pszModule = pReq->u.In.pszModule;
4375	size_t cbSymbol;
4376	char const *pszEnd;
4377	uint32_t i;
4378
4379	/*
4380	* Input validation.
4381	*/
4382	AssertPtrReturn(pszSymbol, VERR_INVALID_POINTER);
4383	pszEnd = RTStrEnd(pszSymbol, 512);
4384	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
4385	cbSymbol = pszEnd - pszSymbol + 1;
4386
4387	if (pszModule)
4388	{
4389	AssertPtrReturn(pszModule, VERR_INVALID_POINTER);
4390	pszEnd = RTStrEnd(pszModule, 64);
4391	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
4392	}
4393	Log3(("supdrvIDC_LdrGetSymbol: pszModule=%p:{%s} pszSymbol=%p:{%s}\n", pszModule, pszModule, pszSymbol, pszSymbol));
4394
4395
4396	if ( !pszModule
4397	\|\| !strcmp(pszModule, "SupDrv"))
4398	{
4399	/*
4400	* Search the support driver export table.
4401	*/
4402	for (i = 0; i < RT_ELEMENTS(g_aFunctions); i++)
4403	if (!strcmp(g_aFunctions[i].szName, pszSymbol))
4404	{
4405	pReq->u.Out.pfnSymbol = g_aFunctions[i].pfn;
4406	break;
4407	}
4408	}
4409	else
4410	{
4411	/*
4412	* Find the loader image.
4413	*/
4414	PSUPDRVLDRIMAGE pImage;
4415
4416	supdrvLdrLock(pDevExt);
4417
4418	for (pImage = pDevExt->pLdrImages; pImage; pImage = pImage->pNext)
4419	if (!strcmp(pImage->szName, pszModule))
4420	break;
4421	if (pImage && pImage->uState == SUP_IOCTL_LDR_LOAD)
4422	{
4423	/*
4424	* Search the symbol strings.
4425	*/
4426	const char *pchStrings = pImage->pachStrTab;
4427	PCSUPLDRSYM paSyms = pImage->paSymbols;
4428	for (i = 0; i < pImage->cSymbols; i++)
4429	{
4430	if ( paSyms[i].offName + cbSymbol <= pImage->cbStrTab
4431	&& !memcmp(pchStrings + paSyms[i].offName, pszSymbol, cbSymbol))
4432	{
4433	/*
4434	* Found it! Calc the symbol address and add a reference to the module.
4435	*/
4436	pReq->u.Out.pfnSymbol = (PFNRT)((uint8_t *)pImage->pvImage + (int32_t)paSyms[i].offSymbol);
4437	rc = supdrvLdrAddUsage(pSession, pImage);
4438	break;
4439	}
4440	}
4441	}
4442	else
4443	rc = pImage ? VERR_WRONG_ORDER : VERR_MODULE_NOT_FOUND;
4444
4445	supdrvLdrUnlock(pDevExt);
4446	}
4447	return rc;
4448	}
4449
4450
4451	/**
4452	* Updates the VMMR0 entry point pointers.
4453	*
4454	* @returns IPRT status code.
4455	* @param pDevExt Device globals.
4456	* @param pSession Session data.
4457	* @param pVMMR0 VMMR0 image handle.
4458	* @param pvVMMR0EntryInt VMMR0EntryInt address.
4459	* @param pvVMMR0EntryFast VMMR0EntryFast address.
4460	* @param pvVMMR0EntryEx VMMR0EntryEx address.
4461	* @remark Caller must own the loader mutex.
4462	*/
4463	static int supdrvLdrSetVMMR0EPs(PSUPDRVDEVEXT pDevExt, void pvVMMR0, void pvVMMR0EntryInt, void pvVMMR0EntryFast, void pvVMMR0EntryEx)
4464	{
4465	int rc = VINF_SUCCESS;
4466	LogFlow(("supdrvLdrSetR0EP pvVMMR0=%p pvVMMR0EntryInt=%p\n", pvVMMR0, pvVMMR0EntryInt));
4467
4468
4469	/*
4470	* Check if not yet set.
4471	*/
4472	if (!pDevExt->pvVMMR0)
4473	{
4474	pDevExt->pvVMMR0 = pvVMMR0;
4475	pDevExt->pfnVMMR0EntryInt = pvVMMR0EntryInt;
4476	pDevExt->pfnVMMR0EntryFast = pvVMMR0EntryFast;
4477	pDevExt->pfnVMMR0EntryEx = pvVMMR0EntryEx;
4478	}
4479	else
4480	{
4481	/*
4482	* Return failure or success depending on whether the values match or not.
4483	*/
4484	if ( pDevExt->pvVMMR0 != pvVMMR0
4485	\|\| (void *)pDevExt->pfnVMMR0EntryInt != pvVMMR0EntryInt
4486	\|\| (void *)pDevExt->pfnVMMR0EntryFast != pvVMMR0EntryFast
4487	\|\| (void *)pDevExt->pfnVMMR0EntryEx != pvVMMR0EntryEx)
4488	{
4489	AssertMsgFailed(("SUP_IOCTL_LDR_SETR0EP: Already set pointing to a different module!\n"));
4490	rc = VERR_INVALID_PARAMETER;
4491	}
4492	}
4493	return rc;
4494	}
4495
4496
4497	/**
4498	* Unsets the VMMR0 entry point installed by supdrvLdrSetR0EP.
4499	*
4500	* @param pDevExt Device globals.
4501	*/
4502	static void supdrvLdrUnsetVMMR0EPs(PSUPDRVDEVEXT pDevExt)
4503	{
4504	pDevExt->pvVMMR0 = NULL;
4505	pDevExt->pfnVMMR0EntryInt = NULL;
4506	pDevExt->pfnVMMR0EntryFast = NULL;
4507	pDevExt->pfnVMMR0EntryEx = NULL;
4508	}
4509
4510
4511	/**
4512	* Adds a usage reference in the specified session of an image.
4513	*
4514	* Called while owning the loader semaphore.
4515	*
4516	* @returns VINF_SUCCESS on success and VERR_NO_MEMORY on failure.
4517	* @param pSession Session in question.
4518	* @param pImage Image which the session is using.
4519	*/
4520	static int supdrvLdrAddUsage(PSUPDRVSESSION pSession, PSUPDRVLDRIMAGE pImage)
4521	{
4522	PSUPDRVLDRUSAGE pUsage;
4523	LogFlow(("supdrvLdrAddUsage: pImage=%p\n", pImage));
4524
4525	/*
4526	* Referenced it already?
4527	*/
4528	pUsage = pSession->pLdrUsage;
4529	while (pUsage)
4530	{
4531	if (pUsage->pImage == pImage)
4532	{
4533	pUsage->cUsage++;
4534	return VINF_SUCCESS;
4535	}
4536	pUsage = pUsage->pNext;
4537	}
4538
4539	/*
4540	* Allocate new usage record.
4541	*/
4542	pUsage = (PSUPDRVLDRUSAGE)RTMemAlloc(sizeof(*pUsage));
4543	AssertReturn(pUsage, /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_5);
4544	pUsage->cUsage = 1;
4545	pUsage->pImage = pImage;
4546	pUsage->pNext = pSession->pLdrUsage;
4547	pSession->pLdrUsage = pUsage;
4548	return VINF_SUCCESS;
4549	}
4550
4551
4552	/**
4553	* Frees a load image.
4554	*
4555	* @param pDevExt Pointer to device extension.
4556	* @param pImage Pointer to the image we're gonna free.
4557	* This image must exit!
4558	* @remark The caller MUST own SUPDRVDEVEXT::mtxLdr!
4559	*/
4560	static void supdrvLdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVLDRIMAGE pImage)
4561	{
4562	PSUPDRVLDRIMAGE pImagePrev;
4563	LogFlow(("supdrvLdrFree: pImage=%p\n", pImage));
4564
4565	/* find it - arg. should've used doubly linked list. */
4566	Assert(pDevExt->pLdrImages);
4567	pImagePrev = NULL;
4568	if (pDevExt->pLdrImages != pImage)
4569	{
4570	pImagePrev = pDevExt->pLdrImages;
4571	while (pImagePrev->pNext != pImage)
4572	pImagePrev = pImagePrev->pNext;
4573	Assert(pImagePrev->pNext == pImage);
4574	}
4575
4576	/* unlink */
4577	if (pImagePrev)
4578	pImagePrev->pNext = pImage->pNext;
4579	else
4580	pDevExt->pLdrImages = pImage->pNext;
4581
4582	/* check if this is VMMR0.r0 unset its entry point pointers. */
4583	if (pDevExt->pvVMMR0 == pImage->pvImage)
4584	supdrvLdrUnsetVMMR0EPs(pDevExt);
4585
4586	/* check for objects with destructors in this image. (Shouldn't happen.) */
4587	if (pDevExt->pObjs)
4588	{
4589	unsigned cObjs = 0;
4590	PSUPDRVOBJ pObj;
4591	RTSpinlockAcquire(pDevExt->Spinlock);
4592	for (pObj = pDevExt->pObjs; pObj; pObj = pObj->pNext)
4593	if (RT_UNLIKELY((uintptr_t)pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImageBits))
4594	{
4595	pObj->pfnDestructor = NULL;
4596	cObjs++;
4597	}
4598	RTSpinlockRelease(pDevExt->Spinlock);
4599	if (cObjs)
4600	OSDBGPRINT(("supdrvLdrFree: Image '%s' has %d dangling objects!\n", pImage->szName, cObjs));
4601	}
4602
4603	/* call termination function if fully loaded. */
4604	if ( pImage->pfnModuleTerm
4605	&& pImage->uState == SUP_IOCTL_LDR_LOAD)
4606	{
4607	LogFlow(("supdrvIOCtl_LdrLoad: calling pfnModuleTerm=%p\n", pImage->pfnModuleTerm));
4608	pImage->pfnModuleTerm(pImage);
4609	}
4610
4611	/* Inform the tracing component. */
4612	supdrvTracerModuleUnloading(pDevExt, pImage);
4613
4614	/* do native unload if appropriate. */
4615	if (pImage->fNative)
4616	supdrvOSLdrUnload(pDevExt, pImage);
4617
4618	/* free the image */
4619	pImage->cUsage = 0;
4620	pImage->pDevExt = NULL;
4621	pImage->pNext = NULL;
4622	pImage->uState = SUP_IOCTL_LDR_FREE;
4623	RTMemExecFree(pImage->pvImageAlloc, pImage->cbImageBits + 31);
4624	pImage->pvImageAlloc = NULL;
4625	RTMemFree(pImage->pachStrTab);
4626	pImage->pachStrTab = NULL;
4627	RTMemFree(pImage->paSymbols);
4628	pImage->paSymbols = NULL;
4629	RTMemFree(pImage);
4630	}
4631
4632
4633	/**
4634	* Acquires the loader lock.
4635	*
4636	* @returns IPRT status code.
4637	* @param pDevExt The device extension.
4638	*/
4639	DECLINLINE(int) supdrvLdrLock(PSUPDRVDEVEXT pDevExt)
4640	{
4641	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
4642	int rc = RTSemMutexRequest(pDevExt->mtxLdr, RT_INDEFINITE_WAIT);
4643	#else
4644	int rc = RTSemFastMutexRequest(pDevExt->mtxLdr);
4645	#endif
4646	AssertRC(rc);
4647	return rc;
4648	}
4649
4650
4651	/**
4652	* Releases the loader lock.
4653	*
4654	* @returns IPRT status code.
4655	* @param pDevExt The device extension.
4656	*/
4657	DECLINLINE(int) supdrvLdrUnlock(PSUPDRVDEVEXT pDevExt)
4658	{
4659	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
4660	return RTSemMutexRelease(pDevExt->mtxLdr);
4661	#else
4662	return RTSemFastMutexRelease(pDevExt->mtxLdr);
4663	#endif
4664	}
4665
4666
4667	/**
4668	* Implements the service call request.
4669	*
4670	* @returns VBox status code.
4671	* @param pDevExt The device extension.
4672	* @param pSession The calling session.
4673	* @param pReq The request packet, valid.
4674	*/
4675	static int supdrvIOCtl_CallServiceModule(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPCALLSERVICE pReq)
4676	{
4677	#if !defined(RT_OS_WINDOWS) \|\| defined(DEBUG)
4678	int rc;
4679
4680	/*
4681	* Find the module first in the module referenced by the calling session.
4682	*/
4683	rc = supdrvLdrLock(pDevExt);
4684	if (RT_SUCCESS(rc))
4685	{
4686	PFNSUPR0SERVICEREQHANDLER pfnServiceReqHandler = NULL;
4687	PSUPDRVLDRUSAGE pUsage;
4688
4689	for (pUsage = pSession->pLdrUsage; pUsage; pUsage = pUsage->pNext)
4690	if ( pUsage->pImage->pfnServiceReqHandler
4691	&& !strcmp(pUsage->pImage->szName, pReq->u.In.szName))
4692	{
4693	pfnServiceReqHandler = pUsage->pImage->pfnServiceReqHandler;
4694	break;
4695	}
4696	supdrvLdrUnlock(pDevExt);
4697
4698	if (pfnServiceReqHandler)
4699	{
4700	/*
4701	* Call it.
4702	*/
4703	if (pReq->Hdr.cbIn == SUP_IOCTL_CALL_SERVICE_SIZE(0))
4704	rc = pfnServiceReqHandler(pSession, pReq->u.In.uOperation, pReq->u.In.u64Arg, NULL);
4705	else
4706	rc = pfnServiceReqHandler(pSession, pReq->u.In.uOperation, pReq->u.In.u64Arg, (PSUPR0SERVICEREQHDR)&pReq->abReqPkt[0]);
4707	}
4708	else
4709	rc = VERR_SUPDRV_SERVICE_NOT_FOUND;
4710	}
4711
4712	/* log it */
4713	if ( RT_FAILURE(rc)
4714	&& rc != VERR_INTERRUPTED
4715	&& rc != VERR_TIMEOUT)
4716	Log(("SUP_IOCTL_CALL_SERVICE: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
4717	rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
4718	else
4719	Log4(("SUP_IOCTL_CALL_SERVICE: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
4720	rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
4721	return rc;
4722	#else /* RT_OS_WINDOWS && !DEBUG */
4723	return VERR_NOT_IMPLEMENTED;
4724	#endif /* RT_OS_WINDOWS && !DEBUG */
4725	}
4726
4727
4728	/**
4729	* Implements the logger settings request.
4730	*
4731	* @returns VBox status code.
4732	* @param pDevExt The device extension.
4733	* @param pSession The caller's session.
4734	* @param pReq The request.
4735	*/
4736	static int supdrvIOCtl_LoggerSettings(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLOGGERSETTINGS pReq)
4737	{
4738	const char *pszGroup = &pReq->u.In.szStrings[pReq->u.In.offGroups];
4739	const char *pszFlags = &pReq->u.In.szStrings[pReq->u.In.offFlags];
4740	const char *pszDest = &pReq->u.In.szStrings[pReq->u.In.offDestination];
4741	PRTLOGGER pLogger = NULL;
4742	int rc;
4743
4744	/*
4745	* Some further validation.
4746	*/
4747	switch (pReq->u.In.fWhat)
4748	{
4749	case SUPLOGGERSETTINGS_WHAT_SETTINGS:
4750	case SUPLOGGERSETTINGS_WHAT_CREATE:
4751	break;
4752
4753	case SUPLOGGERSETTINGS_WHAT_DESTROY:
4754	if (pszGroup \|\| pszFlags \|\| *pszDest)
4755	return VERR_INVALID_PARAMETER;
4756	if (pReq->u.In.fWhich == SUPLOGGERSETTINGS_WHICH_RELEASE)
4757	return VERR_ACCESS_DENIED;
4758	break;
4759
4760	default:
4761	return VERR_INTERNAL_ERROR;
4762	}
4763
4764	/*
4765	* Get the logger.
4766	*/
4767	switch (pReq->u.In.fWhich)
4768	{
4769	case SUPLOGGERSETTINGS_WHICH_DEBUG:
4770	pLogger = RTLogGetDefaultInstance();
4771	break;
4772
4773	case SUPLOGGERSETTINGS_WHICH_RELEASE:
4774	pLogger = RTLogRelDefaultInstance();
4775	break;
4776
4777	default:
4778	return VERR_INTERNAL_ERROR;
4779	}
4780
4781	/*
4782	* Do the job.
4783	*/
4784	switch (pReq->u.In.fWhat)
4785	{
4786	case SUPLOGGERSETTINGS_WHAT_SETTINGS:
4787	if (pLogger)
4788	{
4789	rc = RTLogFlags(pLogger, pszFlags);
4790	if (RT_SUCCESS(rc))
4791	rc = RTLogGroupSettings(pLogger, pszGroup);
4792	NOREF(pszDest);
4793	}
4794	else
4795	rc = VERR_NOT_FOUND;
4796	break;
4797
4798	case SUPLOGGERSETTINGS_WHAT_CREATE:
4799	{
4800	if (pLogger)
4801	rc = VERR_ALREADY_EXISTS;
4802	else
4803	{
4804	static const char * const s_apszGroups[] = VBOX_LOGGROUP_NAMES;
4805
4806	rc = RTLogCreate(&pLogger,
4807	0 /* fFlags */,
4808	pszGroup,
4809	pReq->u.In.fWhich == SUPLOGGERSETTINGS_WHICH_DEBUG
4810	? "VBOX_LOG"
4811	: "VBOX_RELEASE_LOG",
4812	RT_ELEMENTS(s_apszGroups),
4813	s_apszGroups,
4814	RTLOGDEST_STDOUT \| RTLOGDEST_DEBUGGER,
4815	NULL);
4816	if (RT_SUCCESS(rc))
4817	{
4818	rc = RTLogFlags(pLogger, pszFlags);
4819	NOREF(pszDest);
4820	if (RT_SUCCESS(rc))
4821	{
4822	switch (pReq->u.In.fWhich)
4823	{
4824	case SUPLOGGERSETTINGS_WHICH_DEBUG:
4825	pLogger = RTLogSetDefaultInstance(pLogger);
4826	break;
4827	case SUPLOGGERSETTINGS_WHICH_RELEASE:
4828	pLogger = RTLogRelSetDefaultInstance(pLogger);
4829	break;
4830	}
4831	}
4832	RTLogDestroy(pLogger);
4833	}
4834	}
4835	break;
4836	}
4837
4838	case SUPLOGGERSETTINGS_WHAT_DESTROY:
4839	switch (pReq->u.In.fWhich)
4840	{
4841	case SUPLOGGERSETTINGS_WHICH_DEBUG:
4842	pLogger = RTLogSetDefaultInstance(NULL);
4843	break;
4844	case SUPLOGGERSETTINGS_WHICH_RELEASE:
4845	pLogger = RTLogRelSetDefaultInstance(NULL);
4846	break;
4847	}
4848	rc = RTLogDestroy(pLogger);
4849	break;
4850
4851	default:
4852	{
4853	rc = VERR_INTERNAL_ERROR;
4854	break;
4855	}
4856	}
4857
4858	return rc;
4859	}
4860
4861
4862	/**
4863	* Creates the GIP.
4864	*
4865	* @returns VBox status code.
4866	* @param pDevExt Instance data. GIP stuff may be updated.
4867	*/
4868	static int supdrvGipCreate(PSUPDRVDEVEXT pDevExt)
4869	{
4870	PSUPGLOBALINFOPAGE pGip;
4871	RTHCPHYS HCPhysGip;
4872	uint32_t u32SystemResolution;
4873	uint32_t u32Interval;
4874	unsigned cCpus;
4875	int rc;
4876
4877
4878	LogFlow(("supdrvGipCreate:\n"));
4879
4880	/* assert order */
4881	Assert(pDevExt->u32SystemTimerGranularityGrant == 0);
4882	Assert(pDevExt->GipMemObj == NIL_RTR0MEMOBJ);
4883	Assert(!pDevExt->pGipTimer);
4884
4885	/*
4886	* Check the CPU count.
4887	*/
4888	cCpus = RTMpGetArraySize();
4889	if ( cCpus > RTCPUSET_MAX_CPUS
4890	\|\| cCpus > 256 /ApicId is used for the mappings/)
4891	{
4892	SUPR0Printf("VBoxDrv: Too many CPUs (%u) for the GIP (max %u)\n", cCpus, RT_MIN(RTCPUSET_MAX_CPUS, 256));
4893	return VERR_TOO_MANY_CPUS;
4894	}
4895
4896	/*
4897	* Allocate a contiguous set of pages with a default kernel mapping.
4898	*/
4899	rc = RTR0MemObjAllocCont(&pDevExt->GipMemObj, RT_UOFFSETOF(SUPGLOBALINFOPAGE, aCPUs[cCpus]), false /fExecutable/);
4900	if (RT_FAILURE(rc))
4901	{
4902	OSDBGPRINT(("supdrvGipCreate: failed to allocate the GIP page. rc=%d\n", rc));
4903	return rc;
4904	}
4905	pGip = (PSUPGLOBALINFOPAGE)RTR0MemObjAddress(pDevExt->GipMemObj); AssertPtr(pGip);
4906	HCPhysGip = RTR0MemObjGetPagePhysAddr(pDevExt->GipMemObj, 0); Assert(HCPhysGip != NIL_RTHCPHYS);
4907
4908	/*
4909	* Find a reasonable update interval and initialize the structure.
4910	*/
4911	u32Interval = u32SystemResolution = RTTimerGetSystemGranularity();
4912	while (u32Interval < 10000000 /* 10 ms */)
4913	u32Interval += u32SystemResolution;
4914
4915	supdrvGipInit(pDevExt, pGip, HCPhysGip, RTTimeSystemNanoTS(), 1000000000 / u32Interval /=Hz/, cCpus);
4916
4917	/*
4918	* Create the timer.
4919	* If CPU_ALL isn't supported we'll have to fall back to synchronous mode.
4920	*/
4921	if (pGip->u32Mode == SUPGIPMODE_ASYNC_TSC)
4922	{
4923	rc = RTTimerCreateEx(&pDevExt->pGipTimer, u32Interval, RTTIMER_FLAGS_CPU_ALL, supdrvGipAsyncTimer, pDevExt);
4924	if (rc == VERR_NOT_SUPPORTED)
4925	{
4926	OSDBGPRINT(("supdrvGipCreate: omni timer not supported, falling back to synchronous mode\n"));
4927	pGip->u32Mode = SUPGIPMODE_SYNC_TSC;
4928	}
4929	}
4930	if (pGip->u32Mode != SUPGIPMODE_ASYNC_TSC)
4931	rc = RTTimerCreateEx(&pDevExt->pGipTimer, u32Interval, 0, supdrvGipSyncTimer, pDevExt);
4932	if (RT_SUCCESS(rc))
4933	{
4934	rc = RTMpNotificationRegister(supdrvGipMpEvent, pDevExt);
4935	if (RT_SUCCESS(rc))
4936	{
4937	rc = RTMpOnAll(supdrvGipInitOnCpu, pDevExt, pGip);
4938	if (RT_SUCCESS(rc))
4939	{
4940	/*
4941	* We're good.
4942	*/
4943	Log(("supdrvGipCreate: %u ns interval.\n", u32Interval));
4944	g_pSUPGlobalInfoPage = pGip;
4945	return VINF_SUCCESS;
4946	}
4947
4948	OSDBGPRINT(("supdrvGipCreate: RTMpOnAll failed with rc=%Rrc\n", rc));
4949	RTMpNotificationDeregister(supdrvGipMpEvent, pDevExt);
4950
4951	}
4952	else
4953	OSDBGPRINT(("supdrvGipCreate: failed to register MP event notfication. rc=%Rrc\n", rc));
4954	}
4955	else
4956	{
4957	OSDBGPRINT(("supdrvGipCreate: failed create GIP timer at %u ns interval. rc=%Rrc\n", u32Interval, rc));
4958	Assert(!pDevExt->pGipTimer);
4959	}
4960	supdrvGipDestroy(pDevExt);
4961	return rc;
4962	}
4963
4964
4965	/**
4966	* Terminates the GIP.
4967	*
4968	* @param pDevExt Instance data. GIP stuff may be updated.
4969	*/
4970	static void supdrvGipDestroy(PSUPDRVDEVEXT pDevExt)
4971	{
4972	int rc;
4973	#ifdef DEBUG_DARWIN_GIP
4974	OSDBGPRINT(("supdrvGipDestroy: pDevExt=%p pGip=%p pGipTimer=%p GipMemObj=%p\n", pDevExt,
4975	pDevExt->GipMemObj != NIL_RTR0MEMOBJ ? RTR0MemObjAddress(pDevExt->GipMemObj) : NULL,
4976	pDevExt->pGipTimer, pDevExt->GipMemObj));
4977	#endif
4978
4979	/*
4980	* Invalid the GIP data.
4981	*/
4982	if (pDevExt->pGip)
4983	{
4984	supdrvGipTerm(pDevExt->pGip);
4985	pDevExt->pGip = NULL;
4986	}
4987	g_pSUPGlobalInfoPage = NULL;
4988
4989	/*
4990	* Destroy the timer and free the GIP memory object.
4991	*/
4992	if (pDevExt->pGipTimer)
4993	{
4994	rc = RTTimerDestroy(pDevExt->pGipTimer); AssertRC(rc);
4995	pDevExt->pGipTimer = NULL;
4996	}
4997
4998	if (pDevExt->GipMemObj != NIL_RTR0MEMOBJ)
4999	{
5000	rc = RTR0MemObjFree(pDevExt->GipMemObj, true /* free mappings */); AssertRC(rc);
5001	pDevExt->GipMemObj = NIL_RTR0MEMOBJ;
5002	}
5003
5004	/*
5005	* Finally, make sure we've release the system timer resolution request
5006	* if one actually succeeded and is still pending.
5007	*/
5008	if (pDevExt->u32SystemTimerGranularityGrant)
5009	{
5010	rc = RTTimerReleaseSystemGranularity(pDevExt->u32SystemTimerGranularityGrant); AssertRC(rc);
5011	pDevExt->u32SystemTimerGranularityGrant = 0;
5012	}
5013	}
5014
5015
5016	/**
5017	* Timer callback function sync GIP mode.
5018	* @param pTimer The timer.
5019	* @param pvUser The device extension.
5020	*/
5021	static DECLCALLBACK(void) supdrvGipSyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick)
5022	{
5023	RTCCUINTREG fOldFlags = ASMIntDisableFlags(); /* No interruptions please (real problem on S10). */
5024	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
5025	uint64_t u64TSC = ASMReadTSC();
5026	uint64_t NanoTS = RTTimeSystemNanoTS();
5027
5028	supdrvGipUpdate(pDevExt, NanoTS, u64TSC, NIL_RTCPUID, iTick);
5029
5030	ASMSetFlags(fOldFlags);
5031	}
5032
5033
5034	/**
5035	* Timer callback function for async GIP mode.
5036	* @param pTimer The timer.
5037	* @param pvUser The device extension.
5038	*/
5039	static DECLCALLBACK(void) supdrvGipAsyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick)
5040	{
5041	RTCCUINTREG fOldFlags = ASMIntDisableFlags(); /* No interruptions please (real problem on S10). */
5042	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
5043	RTCPUID idCpu = RTMpCpuId();
5044	uint64_t u64TSC = ASMReadTSC();
5045	uint64_t NanoTS = RTTimeSystemNanoTS();
5046
5047	/** @todo reset the transaction number and whatnot when iTick == 1. */
5048	if (pDevExt->idGipMaster == idCpu)
5049	supdrvGipUpdate(pDevExt, NanoTS, u64TSC, idCpu, iTick);
5050	else
5051	supdrvGipUpdatePerCpu(pDevExt, NanoTS, u64TSC, idCpu, ASMGetApicId(), iTick);
5052
5053	ASMSetFlags(fOldFlags);
5054	}
5055
5056
5057	/**
5058	* Finds our (@a idCpu) entry, or allocates a new one if not found.
5059	*
5060	* @returns Index of the CPU in the cache set.
5061	* @param pGip The GIP.
5062	* @param idCpu The CPU ID.
5063	*/
5064	static uint32_t supdrvGipCpuIndexFromCpuId(PSUPGLOBALINFOPAGE pGip, RTCPUID idCpu)
5065	{
5066	uint32_t i, cTries;
5067
5068	/*
5069	* ASSUMES that CPU IDs are constant.
5070	*/
5071	for (i = 0; i < pGip->cCpus; i++)
5072	if (pGip->aCPUs[i].idCpu == idCpu)
5073	return i;
5074
5075	cTries = 0;
5076	do
5077	{
5078	for (i = 0; i < pGip->cCpus; i++)
5079	{
5080	bool fRc;
5081	ASMAtomicCmpXchgSize(&pGip->aCPUs[i].idCpu, idCpu, NIL_RTCPUID, fRc);
5082	if (fRc)
5083	return i;
5084	}
5085	} while (cTries++ < 32);
5086	AssertReleaseFailed();
5087	return i - 1;
5088	}
5089
5090
5091	/**
5092	* The calling CPU should be accounted as online, update GIP accordingly.
5093	*
5094	* This is used by supdrvGipMpEvent as well as the supdrvGipCreate.
5095	*
5096	* @param pDevExt The device extension.
5097	* @param idCpu The CPU ID.
5098	*/
5099	static void supdrvGipMpEventOnline(PSUPDRVDEVEXT pDevExt, RTCPUID idCpu)
5100	{
5101	int iCpuSet = 0;
5102	uint16_t idApic = UINT16_MAX;
5103	uint32_t i = 0;
5104	uint64_t u64NanoTS = 0;
5105	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
5106
5107	AssertPtrReturnVoid(pGip);
5108	AssertRelease(idCpu == RTMpCpuId());
5109	Assert(pGip->cPossibleCpus == RTMpGetCount());
5110
5111	/*
5112	* Do this behind a spinlock with interrupts disabled as this can fire
5113	* on all CPUs simultaneously, see #6110.
5114	*/
5115	RTSpinlockAcquire(pDevExt->hGipSpinlock);
5116
5117	/*
5118	* Update the globals.
5119	*/
5120	ASMAtomicWriteU16(&pGip->cPresentCpus, RTMpGetPresentCount());
5121	ASMAtomicWriteU16(&pGip->cOnlineCpus, RTMpGetOnlineCount());
5122	iCpuSet = RTMpCpuIdToSetIndex(idCpu);
5123	if (iCpuSet >= 0)
5124	{
5125	Assert(RTCpuSetIsMemberByIndex(&pGip->PossibleCpuSet, iCpuSet));
5126	RTCpuSetAddByIndex(&pGip->OnlineCpuSet, iCpuSet);
5127	RTCpuSetAddByIndex(&pGip->PresentCpuSet, iCpuSet);
5128	}
5129
5130	/*
5131	* Update the entry.
5132	*/
5133	u64NanoTS = RTTimeSystemNanoTS() - pGip->u32UpdateIntervalNS;
5134	i = supdrvGipCpuIndexFromCpuId(pGip, idCpu);
5135	supdrvGipInitCpu(pGip, &pGip->aCPUs[i], u64NanoTS);
5136	idApic = ASMGetApicId();
5137	ASMAtomicWriteU16(&pGip->aCPUs[i].idApic, idApic);
5138	ASMAtomicWriteS16(&pGip->aCPUs[i].iCpuSet, (int16_t)iCpuSet);
5139	ASMAtomicWriteSize(&pGip->aCPUs[i].idCpu, idCpu);
5140
5141	/*
5142	* Update the APIC ID and CPU set index mappings.
5143	*/
5144	ASMAtomicWriteU16(&pGip->aiCpuFromApicId[idApic], i);
5145	ASMAtomicWriteU16(&pGip->aiCpuFromCpuSetIdx[iCpuSet], i);
5146
5147	/* commit it */
5148	ASMAtomicWriteSize(&pGip->aCPUs[i].enmState, SUPGIPCPUSTATE_ONLINE);
5149
5150	RTSpinlockReleaseNoInts(pDevExt->hGipSpinlock);
5151	}
5152
5153
5154	/**
5155	* The CPU should be accounted as offline, update the GIP accordingly.
5156	*
5157	* This is used by supdrvGipMpEvent.
5158	*
5159	* @param pDevExt The device extension.
5160	* @param idCpu The CPU ID.
5161	*/
5162	static void supdrvGipMpEventOffline(PSUPDRVDEVEXT pDevExt, RTCPUID idCpu)
5163	{
5164	int iCpuSet;
5165	unsigned i;
5166
5167	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
5168
5169	AssertPtrReturnVoid(pGip);
5170	RTSpinlockAcquire(pDevExt->hGipSpinlock);
5171
5172	iCpuSet = RTMpCpuIdToSetIndex(idCpu);
5173	AssertReturnVoid(iCpuSet >= 0);
5174
5175	i = pGip->aiCpuFromCpuSetIdx[iCpuSet];
5176	AssertReturnVoid(i < pGip->cCpus);
5177	AssertReturnVoid(pGip->aCPUs[i].idCpu == idCpu);
5178
5179	Assert(RTCpuSetIsMemberByIndex(&pGip->PossibleCpuSet, iCpuSet));
5180	RTCpuSetDelByIndex(&pGip->OnlineCpuSet, iCpuSet);
5181
5182	/* commit it */
5183	ASMAtomicWriteSize(&pGip->aCPUs[i].enmState, SUPGIPCPUSTATE_OFFLINE);
5184
5185	RTSpinlockReleaseNoInts(pDevExt->hGipSpinlock);
5186	}
5187
5188
5189	/**
5190	* Multiprocessor event notification callback.
5191	*
5192	* This is used to make sure that the GIP master gets passed on to
5193	* another CPU. It also updates the associated CPU data.
5194	*
5195	* @param enmEvent The event.
5196	* @param idCpu The cpu it applies to.
5197	* @param pvUser Pointer to the device extension.
5198	*
5199	* @remarks This function -must- fire on the newly online'd CPU for the
5200	* RTMPEVENT_ONLINE case and can fire on any CPU for the
5201	* RTMPEVENT_OFFLINE case.
5202	*/
5203	static DECLCALLBACK(void) supdrvGipMpEvent(RTMPEVENT enmEvent, RTCPUID idCpu, void *pvUser)
5204	{
5205	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
5206	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
5207
5208	AssertRelease(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
5209
5210	/*
5211	* Update the GIP CPU data.
5212	*/
5213	if (pGip)
5214	{
5215	switch (enmEvent)
5216	{
5217	case RTMPEVENT_ONLINE:
5218	AssertRelease(idCpu == RTMpCpuId());
5219	supdrvGipMpEventOnline(pDevExt, idCpu);
5220	break;
5221	case RTMPEVENT_OFFLINE:
5222	supdrvGipMpEventOffline(pDevExt, idCpu);
5223	break;
5224
5225	}
5226	}
5227
5228	/*
5229	* Make sure there is a master GIP.
5230	*/
5231	if (enmEvent == RTMPEVENT_OFFLINE)
5232	{
5233	RTCPUID idGipMaster = ASMAtomicReadU32(&pDevExt->idGipMaster);
5234	if (idGipMaster == idCpu)
5235	{
5236	/*
5237	* Find a new GIP master.
5238	*/
5239	bool fIgnored;
5240	unsigned i;
5241	RTCPUID idNewGipMaster = NIL_RTCPUID;
5242	RTCPUSET OnlineCpus;
5243	RTMpGetOnlineSet(&OnlineCpus);
5244
5245	for (i = 0; i < RTCPUSET_MAX_CPUS; i++)
5246	{
5247	RTCPUID idCurCpu = RTMpCpuIdFromSetIndex(i);
5248	if ( RTCpuSetIsMember(&OnlineCpus, idCurCpu)
5249	&& idCurCpu != idGipMaster)
5250	{
5251	idNewGipMaster = idCurCpu;
5252	break;
5253	}
5254	}
5255
5256	Log(("supdrvGipMpEvent: Gip master %#lx -> %#lx\n", (long)idGipMaster, (long)idNewGipMaster));
5257	ASMAtomicCmpXchgSize(&pDevExt->idGipMaster, idNewGipMaster, idGipMaster, fIgnored);
5258	NOREF(fIgnored);
5259	}
5260	}
5261	}
5262
5263
5264	/**
5265	* Callback used by supdrvDetermineAsyncTSC to read the TSC on a CPU.
5266	*
5267	* @param idCpu Ignored.
5268	* @param pvUser1 Where to put the TSC.
5269	* @param pvUser2 Ignored.
5270	*/
5271	static DECLCALLBACK(void) supdrvDetermineAsyncTscWorker(RTCPUID idCpu, void pvUser1, void pvUser2)
5272	{
5273	#if 1
5274	ASMAtomicWriteU64((uint64_t volatile *)pvUser1, ASMReadTSC());
5275	#else
5276	(uint64_t )pvUser1 = ASMReadTSC();
5277	#endif
5278	}
5279
5280
5281	/**
5282	* Determine if Async GIP mode is required because of TSC drift.
5283	*
5284	* When using the default/normal timer code it is essential that the time stamp counter
5285	* (TSC) runs never backwards, that is, a read operation to the counter should return
5286	* a bigger value than any previous read operation. This is guaranteed by the latest
5287	* AMD CPUs and by newer Intel CPUs which never enter the C2 state (P4). In any other
5288	* case we have to choose the asynchronous timer mode.
5289	*
5290	* @param poffMin Pointer to the determined difference between different cores.
5291	* @return false if the time stamp counters appear to be synchronized, true otherwise.
5292	*/
5293	static bool supdrvDetermineAsyncTsc(uint64_t *poffMin)
5294	{
5295	/*
5296	* Just iterate all the cpus 8 times and make sure that the TSC is
5297	* ever increasing. We don't bother taking TSC rollover into account.
5298	*/
5299	int iEndCpu = RTMpGetArraySize();
5300	int iCpu;
5301	int cLoops = 8;
5302	bool fAsync = false;
5303	int rc = VINF_SUCCESS;
5304	uint64_t offMax = 0;
5305	uint64_t offMin = ~(uint64_t)0;
5306	uint64_t PrevTsc = ASMReadTSC();
5307
5308	while (cLoops-- > 0)
5309	{
5310	for (iCpu = 0; iCpu < iEndCpu; iCpu++)
5311	{
5312	uint64_t CurTsc;
5313	rc = RTMpOnSpecific(RTMpCpuIdFromSetIndex(iCpu), supdrvDetermineAsyncTscWorker, &CurTsc, NULL);
5314	if (RT_SUCCESS(rc))
5315	{
5316	if (CurTsc <= PrevTsc)
5317	{
5318	fAsync = true;
5319	offMin = offMax = PrevTsc - CurTsc;
5320	Log(("supdrvDetermineAsyncTsc: iCpu=%d cLoops=%d CurTsc=%llx PrevTsc=%llx\n",
5321	iCpu, cLoops, CurTsc, PrevTsc));
5322	break;
5323	}
5324
5325	/* Gather statistics (except the first time). */
5326	if (iCpu != 0 \|\| cLoops != 7)
5327	{
5328	uint64_t off = CurTsc - PrevTsc;
5329	if (off < offMin)
5330	offMin = off;
5331	if (off > offMax)
5332	offMax = off;
5333	Log2(("%d/%d: off=%llx\n", cLoops, iCpu, off));
5334	}
5335
5336	/* Next */
5337	PrevTsc = CurTsc;
5338	}
5339	else if (rc == VERR_NOT_SUPPORTED)
5340	break;
5341	else
5342	AssertMsg(rc == VERR_CPU_NOT_FOUND \|\| rc == VERR_CPU_OFFLINE, ("%d\n", rc));
5343	}
5344
5345	/* broke out of the loop. */
5346	if (iCpu < iEndCpu)
5347	break;
5348	}
5349
5350	poffMin = offMin; / Almost RTMpOnSpecific profiling. */
5351	Log(("supdrvDetermineAsyncTsc: returns %d; iEndCpu=%d rc=%d offMin=%llx offMax=%llx\n",
5352	fAsync, iEndCpu, rc, offMin, offMax));
5353	#if !defined(RT_OS_SOLARIS) && !defined(RT_OS_OS2) && !defined(RT_OS_WINDOWS)
5354	OSDBGPRINT(("vboxdrv: fAsync=%d offMin=%#lx offMax=%#lx\n", fAsync, (long)offMin, (long)offMax));
5355	#endif
5356	return fAsync;
5357	}
5358
5359
5360	/**
5361	* Determine the GIP TSC mode.
5362	*
5363	* @returns The most suitable TSC mode.
5364	* @param pDevExt Pointer to the device instance data.
5365	*/
5366	static SUPGIPMODE supdrvGipDeterminTscMode(PSUPDRVDEVEXT pDevExt)
5367	{
5368	/*
5369	* On SMP we're faced with two problems:
5370	* (1) There might be a skew between the CPU, so that cpu0
5371	* returns a TSC that is slightly different from cpu1.
5372	* (2) Power management (and other things) may cause the TSC
5373	* to run at a non-constant speed, and cause the speed
5374	* to be different on the cpus. This will result in (1).
5375	*
5376	* So, on SMP systems we'll have to select the ASYNC update method
5377	* if there are symptoms of these problems.
5378	*/
5379	if (RTMpGetCount() > 1)
5380	{
5381	uint32_t uEAX, uEBX, uECX, uEDX;
5382	uint64_t u64DiffCoresIgnored;
5383
5384	/* Permit the user and/or the OS specific bits to force async mode. */
5385	if (supdrvOSGetForcedAsyncTscMode(pDevExt))
5386	return SUPGIPMODE_ASYNC_TSC;
5387
5388	/* Try check for current differences between the cpus. */
5389	if (supdrvDetermineAsyncTsc(&u64DiffCoresIgnored))
5390	return SUPGIPMODE_ASYNC_TSC;
5391
5392	/*
5393	* If the CPU supports power management and is an AMD one we
5394	* won't trust it unless it has the TscInvariant bit is set.
5395	*/
5396	/* Check for "AuthenticAMD" */
5397	ASMCpuId(0, &uEAX, &uEBX, &uECX, &uEDX);
5398	if ( uEAX >= 1
5399	&& uEBX == X86_CPUID_VENDOR_AMD_EBX
5400	&& uECX == X86_CPUID_VENDOR_AMD_ECX
5401	&& uEDX == X86_CPUID_VENDOR_AMD_EDX)
5402	{
5403	/* Check for APM support and that TscInvariant is cleared. */
5404	ASMCpuId(0x80000000, &uEAX, &uEBX, &uECX, &uEDX);
5405	if (uEAX >= 0x80000007)
5406	{
5407	ASMCpuId(0x80000007, &uEAX, &uEBX, &uECX, &uEDX);
5408	if ( !(uEDX & RT_BIT(8))/* TscInvariant */
5409	&& (uEDX & 0x3e)) /* STC\|TM\|THERMTRIP\|VID\|FID. Ignore TS. */
5410	return SUPGIPMODE_ASYNC_TSC;
5411	}
5412	}
5413	}
5414	return SUPGIPMODE_SYNC_TSC;
5415	}
5416
5417
5418	/**
5419	* Initializes per-CPU GIP information.
5420	*
5421	* @param pGip Pointer to the read-write kernel mapping of the GIP.
5422	* @param pCpu Pointer to which GIP CPU to initalize.
5423	* @param u64NanoTS The current nanosecond timestamp.
5424	*/
5425	static void supdrvGipInitCpu(PSUPGLOBALINFOPAGE pGip, PSUPGIPCPU pCpu, uint64_t u64NanoTS)
5426	{
5427	pCpu->u32TransactionId = 2;
5428	pCpu->u64NanoTS = u64NanoTS;
5429	pCpu->u64TSC = ASMReadTSC();
5430
5431	ASMAtomicWriteSize(&pCpu->enmState, SUPGIPCPUSTATE_INVALID);
5432	ASMAtomicWriteSize(&pCpu->idCpu, NIL_RTCPUID);
5433	ASMAtomicWriteS16(&pCpu->iCpuSet, -1);
5434	ASMAtomicWriteU16(&pCpu->idApic, UINT16_MAX);
5435
5436	/*
5437	* We don't know the following values until we've executed updates.
5438	* So, we'll just pretend it's a 4 GHz CPU and adjust the history it on
5439	* the 2nd timer callout.
5440	*/
5441	pCpu->u64CpuHz = _4G + 1; /* tstGIP-2 depends on this. */
5442	pCpu->u32UpdateIntervalTSC
5443	= pCpu->au32TSCHistory[0]
5444	= pCpu->au32TSCHistory[1]
5445	= pCpu->au32TSCHistory[2]
5446	= pCpu->au32TSCHistory[3]
5447	= pCpu->au32TSCHistory[4]
5448	= pCpu->au32TSCHistory[5]
5449	= pCpu->au32TSCHistory[6]
5450	= pCpu->au32TSCHistory[7]
5451	= (uint32_t)(_4G / pGip->u32UpdateHz);
5452	}
5453
5454
5455	/**
5456	* Initializes the GIP data.
5457	*
5458	* @param pDevExt Pointer to the device instance data.
5459	* @param pGip Pointer to the read-write kernel mapping of the GIP.
5460	* @param HCPhys The physical address of the GIP.
5461	* @param u64NanoTS The current nanosecond timestamp.
5462	* @param uUpdateHz The update frequency.
5463	* @param cCpus The CPU count.
5464	*/
5465	static void supdrvGipInit(PSUPDRVDEVEXT pDevExt, PSUPGLOBALINFOPAGE pGip, RTHCPHYS HCPhys,
5466	uint64_t u64NanoTS, unsigned uUpdateHz, unsigned cCpus)
5467	{
5468	size_t const cbGip = RT_ALIGN_Z(RT_OFFSETOF(SUPGLOBALINFOPAGE, aCPUs[cCpus]), PAGE_SIZE);
5469	unsigned i;
5470	#ifdef DEBUG_DARWIN_GIP
5471	OSDBGPRINT(("supdrvGipInit: pGip=%p HCPhys=%lx u64NanoTS=%llu uUpdateHz=%d cCpus=%u\n", pGip, (long)HCPhys, u64NanoTS, uUpdateHz, cCpus));
5472	#else
5473	LogFlow(("supdrvGipInit: pGip=%p HCPhys=%lx u64NanoTS=%llu uUpdateHz=%d cCpus=%u\n", pGip, (long)HCPhys, u64NanoTS, uUpdateHz, cCpus));
5474	#endif
5475
5476	/*
5477	* Initialize the structure.
5478	*/
5479	memset(pGip, 0, cbGip);
5480	pGip->u32Magic = SUPGLOBALINFOPAGE_MAGIC;
5481	pGip->u32Version = SUPGLOBALINFOPAGE_VERSION;
5482	pGip->u32Mode = supdrvGipDeterminTscMode(pDevExt);
5483	pGip->cCpus = (uint16_t)cCpus;
5484	pGip->cPages = (uint16_t)(cbGip / PAGE_SIZE);
5485	pGip->u32UpdateHz = uUpdateHz;
5486	pGip->u32UpdateIntervalNS = 1000000000 / uUpdateHz;
5487	pGip->u64NanoTSLastUpdateHz = u64NanoTS;
5488	RTCpuSetEmpty(&pGip->OnlineCpuSet);
5489	RTCpuSetEmpty(&pGip->PresentCpuSet);
5490	RTMpGetSet(&pGip->PossibleCpuSet);
5491	pGip->cOnlineCpus = RTMpGetOnlineCount();
5492	pGip->cPresentCpus = RTMpGetPresentCount();
5493	pGip->cPossibleCpus = RTMpGetCount();
5494	pGip->idCpuMax = RTMpGetMaxCpuId();
5495	for (i = 0; i < RT_ELEMENTS(pGip->aiCpuFromApicId); i++)
5496	pGip->aiCpuFromApicId[i] = 0;
5497	for (i = 0; i < RT_ELEMENTS(pGip->aiCpuFromCpuSetIdx); i++)
5498	pGip->aiCpuFromCpuSetIdx[i] = UINT16_MAX;
5499
5500	for (i = 0; i < cCpus; i++)
5501	supdrvGipInitCpu(pGip, &pGip->aCPUs[i], u64NanoTS);
5502
5503	/*
5504	* Link it to the device extension.
5505	*/
5506	pDevExt->pGip = pGip;
5507	pDevExt->HCPhysGip = HCPhys;
5508	pDevExt->cGipUsers = 0;
5509	}
5510
5511
5512	/**
5513	* On CPU initialization callback for RTMpOnAll.
5514	*
5515	* @param idCpu The CPU ID.
5516	* @param pvUser1 The device extension.
5517	* @param pvUser2 The GIP.
5518	*/
5519	static DECLCALLBACK(void) supdrvGipInitOnCpu(RTCPUID idCpu, void pvUser1, void pvUser2)
5520	{
5521	/* This is good enough, even though it will update some of the globals a
5522	bit to much. */
5523	supdrvGipMpEventOnline((PSUPDRVDEVEXT)pvUser1, idCpu);
5524	}
5525
5526
5527	/**
5528	* Invalidates the GIP data upon termination.
5529	*
5530	* @param pGip Pointer to the read-write kernel mapping of the GIP.
5531	*/
5532	static void supdrvGipTerm(PSUPGLOBALINFOPAGE pGip)
5533	{
5534	unsigned i;
5535	pGip->u32Magic = 0;
5536	for (i = 0; i < RT_ELEMENTS(pGip->aCPUs); i++)
5537	{
5538	pGip->aCPUs[i].u64NanoTS = 0;
5539	pGip->aCPUs[i].u64TSC = 0;
5540	pGip->aCPUs[i].iTSCHistoryHead = 0;
5541	}
5542	}
5543
5544
5545	/**
5546	* Worker routine for supdrvGipUpdate and supdrvGipUpdatePerCpu that
5547	* updates all the per cpu data except the transaction id.
5548	*
5549	* @param pDevExt The device extension.
5550	* @param pGipCpu Pointer to the per cpu data.
5551	* @param u64NanoTS The current time stamp.
5552	* @param u64TSC The current TSC.
5553	* @param iTick The current timer tick.
5554	*/
5555	static void supdrvGipDoUpdateCpu(PSUPDRVDEVEXT pDevExt, PSUPGIPCPU pGipCpu, uint64_t u64NanoTS, uint64_t u64TSC, uint64_t iTick)
5556	{
5557	uint64_t u64TSCDelta;
5558	uint32_t u32UpdateIntervalTSC;
5559	uint32_t u32UpdateIntervalTSCSlack;
5560	unsigned iTSCHistoryHead;
5561	uint64_t u64CpuHz;
5562	uint32_t u32TransactionId;
5563
5564	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
5565	AssertPtrReturnVoid(pGip);
5566
5567	/* Delta between this and the previous update. */
5568	ASMAtomicUoWriteU32(&pGipCpu->u32PrevUpdateIntervalNS, (uint32_t)(u64NanoTS - pGipCpu->u64NanoTS));
5569
5570	/*
5571	* Update the NanoTS.
5572	*/
5573	ASMAtomicWriteU64(&pGipCpu->u64NanoTS, u64NanoTS);
5574
5575	/*
5576	* Calc TSC delta.
5577	*/
5578	/** @todo validate the NanoTS delta, don't trust the OS to call us when it should... */
5579	u64TSCDelta = u64TSC - pGipCpu->u64TSC;
5580	ASMAtomicWriteU64(&pGipCpu->u64TSC, u64TSC);
5581
5582	if (u64TSCDelta >> 32)
5583	{
5584	u64TSCDelta = pGipCpu->u32UpdateIntervalTSC;
5585	pGipCpu->cErrors++;
5586	}
5587
5588	/*
5589	* On the 2nd and 3rd callout, reset the history with the current TSC
5590	* interval since the values entered by supdrvGipInit are totally off.
5591	* The interval on the 1st callout completely unreliable, the 2nd is a bit
5592	* better, while the 3rd should be most reliable.
5593	*/
5594	u32TransactionId = pGipCpu->u32TransactionId;
5595	if (RT_UNLIKELY( ( u32TransactionId == 5
5596	\|\| u32TransactionId == 7)
5597	&& ( iTick == 2
5598	\|\| iTick == 3) ))
5599	{
5600	unsigned i;
5601	for (i = 0; i < RT_ELEMENTS(pGipCpu->au32TSCHistory); i++)
5602	ASMAtomicUoWriteU32(&pGipCpu->au32TSCHistory[i], (uint32_t)u64TSCDelta);
5603	}
5604
5605	/*
5606	* TSC History.
5607	*/
5608	Assert(RT_ELEMENTS(pGipCpu->au32TSCHistory) == 8);
5609	iTSCHistoryHead = (pGipCpu->iTSCHistoryHead + 1) & 7;
5610	ASMAtomicWriteU32(&pGipCpu->iTSCHistoryHead, iTSCHistoryHead);
5611	ASMAtomicWriteU32(&pGipCpu->au32TSCHistory[iTSCHistoryHead], (uint32_t)u64TSCDelta);
5612
5613	/*
5614	* UpdateIntervalTSC = average of last 8,2,1 intervals depending on update HZ.
5615	*/
5616	if (pGip->u32UpdateHz >= 1000)
5617	{
5618	uint32_t u32;
5619	u32 = pGipCpu->au32TSCHistory[0];
5620	u32 += pGipCpu->au32TSCHistory[1];
5621	u32 += pGipCpu->au32TSCHistory[2];
5622	u32 += pGipCpu->au32TSCHistory[3];
5623	u32 >>= 2;
5624	u32UpdateIntervalTSC = pGipCpu->au32TSCHistory[4];
5625	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[5];
5626	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[6];
5627	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[7];
5628	u32UpdateIntervalTSC >>= 2;
5629	u32UpdateIntervalTSC += u32;
5630	u32UpdateIntervalTSC >>= 1;
5631
5632	/* Value chosen for a 2GHz Athlon64 running linux 2.6.10/11, . */
5633	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 14;
5634	}
5635	else if (pGip->u32UpdateHz >= 90)
5636	{
5637	u32UpdateIntervalTSC = (uint32_t)u64TSCDelta;
5638	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[(iTSCHistoryHead - 1) & 7];
5639	u32UpdateIntervalTSC >>= 1;
5640
5641	/* value chosen on a 2GHz thinkpad running windows */
5642	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 7;
5643	}
5644	else
5645	{
5646	u32UpdateIntervalTSC = (uint32_t)u64TSCDelta;
5647
5648	/* This value hasn't be checked yet.. waiting for OS/2 and 33Hz timers.. :-) */
5649	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 6;
5650	}
5651	ASMAtomicWriteU32(&pGipCpu->u32UpdateIntervalTSC, u32UpdateIntervalTSC + u32UpdateIntervalTSCSlack);
5652
5653	/*
5654	* CpuHz.
5655	*/
5656	u64CpuHz = ASMMult2xU32RetU64(u32UpdateIntervalTSC, pGip->u32UpdateHz);
5657	ASMAtomicWriteU64(&pGipCpu->u64CpuHz, u64CpuHz);
5658	}
5659
5660
5661	/**
5662	* Updates the GIP.
5663	*
5664	* @param pDevExt The device extension.
5665	* @param u64NanoTS The current nanosecond timesamp.
5666	* @param u64TSC The current TSC timesamp.
5667	* @param idCpu The CPU ID.
5668	* @param iTick The current timer tick.
5669	*/
5670	static void supdrvGipUpdate(PSUPDRVDEVEXT pDevExt, uint64_t u64NanoTS, uint64_t u64TSC, RTCPUID idCpu, uint64_t iTick)
5671	{
5672	/*
5673	* Determine the relevant CPU data.
5674	*/
5675	PSUPGIPCPU pGipCpu;
5676	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
5677	AssertPtrReturnVoid(pGip);
5678
5679	if (pGip->u32Mode != SUPGIPMODE_ASYNC_TSC)
5680	pGipCpu = &pGip->aCPUs[0];
5681	else
5682	{
5683	unsigned iCpu = pGip->aiCpuFromApicId[ASMGetApicId()];
5684	if (RT_UNLIKELY(iCpu >= pGip->cCpus))
5685	return;
5686	pGipCpu = &pGip->aCPUs[iCpu];
5687	if (RT_UNLIKELY(pGipCpu->idCpu != idCpu))
5688	return;
5689	}
5690
5691	/*
5692	* Start update transaction.
5693	*/
5694	if (!(ASMAtomicIncU32(&pGipCpu->u32TransactionId) & 1))
5695	{
5696	/* this can happen on win32 if we're taking to long and there are more CPUs around. shouldn't happen though. */
5697	AssertMsgFailed(("Invalid transaction id, %#x, not odd!\n", pGipCpu->u32TransactionId));
5698	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
5699	pGipCpu->cErrors++;
5700	return;
5701	}
5702
5703	/*
5704	* Recalc the update frequency every 0x800th time.
5705	*/
5706	if (!(pGipCpu->u32TransactionId & (GIP_UPDATEHZ_RECALC_FREQ * 2 - 2)))
5707	{
5708	if (pGip->u64NanoTSLastUpdateHz)
5709	{
5710	#ifdef RT_ARCH_AMD64 /** @todo fix 64-bit div here to work on x86 linux. */
5711	uint64_t u64Delta = u64NanoTS - pGip->u64NanoTSLastUpdateHz;
5712	uint32_t u32UpdateHz = (uint32_t)((UINT64_C(1000000000) * GIP_UPDATEHZ_RECALC_FREQ) / u64Delta);
5713	if (u32UpdateHz <= 2000 && u32UpdateHz >= 30)
5714	{
5715	ASMAtomicWriteU32(&pGip->u32UpdateHz, u32UpdateHz);
5716	ASMAtomicWriteU32(&pGip->u32UpdateIntervalNS, 1000000000 / u32UpdateHz);
5717	}
5718	#endif
5719	}
5720	ASMAtomicWriteU64(&pGip->u64NanoTSLastUpdateHz, u64NanoTS);
5721	}
5722
5723	/*
5724	* Update the data.
5725	*/
5726	supdrvGipDoUpdateCpu(pDevExt, pGipCpu, u64NanoTS, u64TSC, iTick);
5727
5728	/*
5729	* Complete transaction.
5730	*/
5731	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
5732	}
5733
5734
5735	/**
5736	* Updates the per cpu GIP data for the calling cpu.
5737	*
5738	* @param pDevExt The device extension.
5739	* @param u64NanoTS The current nanosecond timesamp.
5740	* @param u64TSC The current TSC timesamp.
5741	* @param idCpu The CPU ID.
5742	* @param idApic The APIC id for the CPU index.
5743	* @param iTick The current timer tick.
5744	*/
5745	static void supdrvGipUpdatePerCpu(PSUPDRVDEVEXT pDevExt, uint64_t u64NanoTS, uint64_t u64TSC,
5746	RTCPUID idCpu, uint8_t idApic, uint64_t iTick)
5747	{
5748	uint32_t iCpu;
5749	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
5750
5751	/*
5752	* Avoid a potential race when a CPU online notification doesn't fire on
5753	* the onlined CPU but the tick creeps in before the event notification is
5754	* run.
5755	*/
5756	if (RT_UNLIKELY(iTick == 1))
5757	{
5758	iCpu = supdrvGipCpuIndexFromCpuId(pGip, idCpu);
5759	if (pGip->aCPUs[iCpu].enmState == SUPGIPCPUSTATE_OFFLINE)
5760	supdrvGipMpEventOnline(pDevExt, idCpu);
5761	}
5762
5763	iCpu = pGip->aiCpuFromApicId[idApic];
5764	if (RT_LIKELY(iCpu < pGip->cCpus))
5765	{
5766	PSUPGIPCPU pGipCpu = &pGip->aCPUs[iCpu];
5767	if (pGipCpu->idCpu == idCpu)
5768	{
5769	/*
5770	* Start update transaction.
5771	*/
5772	if (!(ASMAtomicIncU32(&pGipCpu->u32TransactionId) & 1))
5773	{
5774	AssertMsgFailed(("Invalid transaction id, %#x, not odd!\n", pGipCpu->u32TransactionId));
5775	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
5776	pGipCpu->cErrors++;
5777	return;
5778	}
5779
5780	/*
5781	* Update the data.
5782	*/
5783	supdrvGipDoUpdateCpu(pDevExt, pGipCpu, u64NanoTS, u64TSC, iTick);
5784
5785	/*
5786	* Complete transaction.
5787	*/
5788	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
5789	}
5790	}
5791	}
5792

Note: See TracBrowser for help on using the repository browser.

source: vbox/trunk/src/VBox/HostDrivers/Support/SUPDrv.c@ 40904

Download in other formats: